Tuesday, 22 February 2011

Reading list redux

It’s been a while since I posted my recommended reading list, so here’s the 2011 version.

Product Details Code complete 2nd edition
Steve McConnel
Microsoft Press

A practical guide to the craft of programming.

Get it from Amazon.co.uk
Product Details Design patterns – elements of reusable object oriented software.
Gamma, Helm, Johnson and Vlissides
Addison Wesley

The famous gang of four book featuring 20 odd common design patterns.

Get it from Amazon.co.uk
Product Details Patterns of enterprise application architecture
Martin Fowler
Addison Wesley

This book details many software design patterns found in enterprise software by the thoughtworks guru, Martin fowler.

Get it from Amazon.co.uk
Product Details The mythical man month
Frederick P. Brooks
Addison Wesley

This is a collection of essays on software project management that was first published in 1975.It contains the now legendary “Brooks Law”, that adding manpower to a late project only makes it later. The concepts laid out in this book are as valid today as they were 30 years ago.

Get it from Amazon.co.uk
Product Details Peopleware
Tom Demarco and Timothy Lister
Dorset house publishing

This book makes the assertion that most development projects fail because of failures within the team – a humorous and riveting read.

Get it from Amazon.co.uk
Product Details Joel on software
Joel Spolsky

I’ve just finished reading this, and the second series (more Joel on software) and it’s an insightful look at the craft of development, the people who do development and running a technology business. Highly recommended.

Get it from Amazon.co.uk
Product Details More Joel on software
Joel Spolsky

As above, this book builds on the previous one – there’s some new stuff and some more depth on some of the topics and again I highly recommend reading it, but if you have a choice between this and the earlier book, choose the earlier one.

Get it from Amazon.co.uk
Product Details Founders at work
Jessiva Livingston

A great read, stories of a number of high profile start ups that have gone on to great things. The book is insightful and for a geek like me who likes to understand tech history, its a staple.

Get it from Amazon.co.uk
Product Details Coders at work
Peter Seibel

This follows a similar vein to the founders at work but is a series of interviews with some great programmers and engineers. Its good to get in the mind of some of these people, but at times the book can be a little dry.

Get it from Amazon.co.uk
Product Details Accidental Empires
Robert X. Cringely
Penguin books

A cracking look at the history of the technology industry – this book is extremely well written and I couldn’t put it down once I started.

Get it from Amazon.co.uk
Product Details Dreaming in code
Scott Rosenberg
Three Rivers Press

Software is hard – this book follows the work of Mitch Kapor’s Chandler project over the course of 3 years and poses the question – why is good software so hard to make?

Get it from Amazon.co.uk

Anyone any further suggestions?

SDLC – Key to a successful development project

In this week’s SDLC related post, I’m going to try to explain how to structure a successful development project, one that when it goes live with users, puts happy smiles on their faces and, because it addresses their needs, massages away the pains they have been having. I’m talking of course here about bespoke development projects (internal development projects) rather than a software house writing the next version of “Micropple iWordPad 2015 Enterprise Edition SP2” – there are subtle yet important differences between the two types of project including;

  • Software house
    • Low cost/revenue per user head
    • Focus on lower cost, high volume sales requires the product to be best in market
    • High degree of spit and polish – can spend significant money on UI design to ensure product looks “the nuts”
    • Ship dates fluid, dictated by feature completion.
  • Bespoke
    • Is best in market by definition as it addresses the businesses very specific needs.
    • Requires spit and polish, but focus is on cost, usability, functionality and implementation time.
      • A fine balance – lack of spit = user rejection, too much = missed functionality
    • Very high cost per user head.

Spit and polish
There’s one contentious issue there – it almost seems like I’m saying that internal or bespoke software won’t be as high quality as a software product from a vendor – that’s not what I mean – quality must still be high, but in terms of going the extra mile to have the most appealing, beautiful and slick UI, unfortunately it is a practical truth that internal development can’t afford to put the same amount of time and money into this as a software house. Internal projects are faced with a trade off between a high level of spit and polish versus factors such as cost and timescales.

Consider a business that is spending £1.4 million on a custom system to, lets say, manage insurance claims and the user base is 200 claim handlers – then the cost per seat of that software is £7,000. Now consider the same business buying something off the shelf for £250,000 that meets around 70% of their requirements – they now have a cost per seat of £1250. The business understands however that the 30% of missing requirements means it won’t deliver on their key objectives and, because their business is so unique (they archive their claim documentation in shredded format for security and file it with a warehouse on the moon), they need to go down the bespoke route – the off the shelf package won’t cut the mustard.

Now they are spending an extra £5750 per user so the product damn well better be good and meet the moon archiving objectives, so you can’t say internal software doesn’t need to have high quality – it absolutely, positively must, but when faced with the choice of tweaking individual pixels in form layouts versus building feature Z (which will save them another £100k per year) the business will choose feature Z every time.

That is the reason I say internal software doesn’t require the same level of spit and polish – it still needs to be slick, but whereas the software house needs to spend the money to polish their UI to make sales, the internal software will often concentrate on features – those features will still be polished, just not to the same extent.

Ok, so that was a long winded way of saying this post is more applicable to bespoke developments rather than the software house – both should follow the same core principle – listen to the customer, but the bespoke team has one advantage – the customer is sat next to them, in the same building, or at least in the same company.

Engaging the customer
Failure in projects is often down to a lack of engagement with the users – they can feel that the new system is forced upon them without consultation and, to be frank, the users of your shiny new system are the ones with the highest understanding of their domain and often have the best suggestions for improvement. As such, it’s important to engage with the users of your system in a variety of ways to let them have their say in how the system should function. So how do we do this at the various stages of the project?

The kick off meeting
At the start of the project, I usually try to get as many users as possible together in one place and present the project to them. This early presentation is a great time to outline the vision for the project, what we’re hoping to achieve and also to set some expectations with them that you want them to be involved at various times. Describing the system and how initially we think it will work and what problems it will address generates an excitement about the project and gets them thinking about how they want to see the project progress.

Of course, at this point you’ve already done some up front estimating to get the client to agree to spend a million pounds with you and the business will have set some goals for when the system needs to be available so you already have constraints at this point, but it should be trivial to accommodate user’s suggestions and ideas, although it has to be said not all of them – you will always find someone with a whacky idea that has no mileage and so as a leader on a project you’re going to need some diplomacy skills!

At the end of the kick off meeting, everyone should understand the objectives of the project and know they are going to be expected to contribute in coming weeks and months. This is their opportunity to change things for the better and to help build a system that will help them do their jobs productively. Allow time at the end for questions and let people voice their concerns and ideas, but guide them into contributing the bulk of their discussions into workshops.

Iterating and workshops
I’ve mentioned in earlier posts that the detailed functional design work should be done an iteration ahead of the actual construction but to get the detail for this design, I like to use workshop sessions with small groups of users. There’s nothing particularly special about these sessions, you get a maximum of around 6 users together in a room, a whiteboard and someone to record important details.

With everyone gathered, we explain the purpose of the workshop and it’s scope (you might have users that work across different functional areas who are keen to get their ideas across) and ask them to describe their job, the processes they follow, where improvements could be made, and what the new application should look like. Time-box the session to around an hour (people’s attention will start to wane beyond this) and ultimately you should end up with whiteboards full of process charts, low fidelity UI designs and some idea of what data you need to capture and how it relates – this, combined with the output from several workshops, should then feed into the design documentation.


An important thing to note here, a successful workshop starts with just the material I’ve mentioned above – I’ve seen misguided analysts conduct workshops where the starting point has been a visio diagram or presentation of how they think the user works and the processes they follow and then proceed to present this to them rather than asking them to collaborate in the design. The problem with this is, users switch off, they come to the conclusion you already know everything you need to know (or think you do, you fool!) and generally won’t engage in a productive design session – if anything they will tear your presentation to pieces because it’s being pitched to them as the way to work rather than asking them to contribute.

Requirements and documentation
From the workshops and all the useful information gleaned from the users, the analyst will then work with the architect to draw up the functional specifications. These are simple use cases written for the user that describe how the system should operate including;

  • Case properties
    • Who does it.
    • How the case is triggered.
    • What conditions dictate the use case is executed.
    • What the outcome should be.
  • Summary
    • A brief description of the purpose and flow of the use case.
  • Normal path
    • The path through the use case in bullet points
      • What the user does, what controls they interact with, the data that is captured.
      • The rules that are checked (conditional logic that can trigger an alternate path)
  • Alternate paths
    • The alternative paths through the use case raised from conditional logic in the normal path

This document should avoid technical implementation details or any specific technology and rather talk in the language of the user. These documents are then reviewed for accuracy by the users and a refinement cycle ensues until everyone agrees that the document describes how this piece of the system should work. These reviews can be done by simply sending out the document and gathering responses, but I find it productive to take the users from the workshops into a discussion where we quickly review and refine the document together (this often drives out any missed details that can be accommodated rapidly with agreement from the product owner and the users in the room).

Technical documentation
With the functional specification complete and agreed, the architect then takes over and transposes the functional details into a physical specification. I’m a firm believer that development is a creative vocation and so the developers should be left to be creative, but given enough information to be able to implement the application within a set of boundaries. As such, the technical specifications are usually done in the form of;

  • Data or domain object model (depending on your preference of data first or objects first)
  • Low fidelity interface mock ups
  • Implementation notes.

This should be all the developer needs to be able to implement the required functionality – the specifics are up to them provided they are following the boundaries of the overall application architecture and framework – for instance your architecture may prescribe NHibernate as an ORM layer, hydrating and persisting aggregate root domain objects through a repository pattern, surfaced in the UI through MVVM using Prism and so on.

Active development and the show and tell
As the construction iteration starts, the functional and technical documentation is provided to the developers, but this is done in a formal way too. The idea being, as the development manager/architect/technical lead, your job is to not only design the system and solve the complex technical issues, but you are the interface between the business and the technical team, you should speak both languages (and for this reason it’s my opinion that the best leads are socially and business aware programmers themselves and should remain active and hands on with code).

At the start of the iteration, my teams hold an iteration planning meeting. At this meeting all the developers involved in the product gather with the analyst, product owner and you as the leader – we go through in absolute detail all of the functional and technical specifications so that everyone has a deep understanding of the requirements and how we are going to implement it – this often involves re-iterating earlier workshops on whiteboards along with discussing the documentation. The programmers are expected to contribute in detail to the ideas of how the implementation should progress, but also in the process of refining the estimates which they should own and take responsibility for.

This can take the form of planning poker or some of the other fun approaches to planning, but my preference is simply to work with the programmers to split the use case down into 1-4 hour individual tasks and estimate them themselves. This planning session gives us a very detailed view of what is involved in the implementation and allows us to compare this to our original high level estimates, the resources and the velocity we have available for that iteration. Where there is slack, we can pad out with more polishing and where there is overrun we can start to look at cutting nice to have functionality from the scope to make the amount of work for the iteration fit to the velocity we have available.

I should note here that the initial iteration plan would be done from the high level estimates and this means you need a good estimator pulling those high level estimates together. If the estimator isn’t a programmer herself, there could be large differences between the high level and detailed plans and this is going to cause you significant problems. In my last project, over 18 iterations, my margin of error between high and low level estimates was around 2% – if this was say, 20% and was optimistic, you’ve just essentially lost one member of a 5 man team due to your bad estimating.

From the iteration planning meeting (I usually allocate one entire day for this), the entire dev team is up to speed with the requirements, has an idea of how to implement everything, understands what tasks need to be done over the iteration and can get cracking. At this point, it’s tempting to become insular and just get on with the development whilst your users get on with their day job, and that is a mistake – the users need to feel engaged continually and you need them to have sight of the emerging product because in doing so, the users are more likely to realise, very early, when things might be going askew.

This review doesn’t need to be formal, doesn’t need to be meetings and presentations, it’s useful to get a handful of users each week for 10-15 minutes crowded around one of the programmers computers to look at how the system is progressing. These show and tell sessions can be extremely valuable in detecting and fixing those small issues that always crop up in dev projects, and getting visibility of them early.

The final mechanism for engagement is the burn-down. People like to see progress, especially management and the people who’s necks are on the line for delivery on time and on budget. On a small team this can be very simple – the individual tasks identified during iteration planning can be put on post it notes and stuck to a whiteboard grid where rows represent the use case being developed (the stickies go in each row representing the use case they belong to) and the columns represent the state of the item – todo, in progress, for test, done (this is a typical scrum board).

At any time, anyone can take a look at the board and see a good visual indication about where the team is up to during that iteration. Each week, or even daily, the team lead should transpose current amount of work remaining onto a line chart showing the progress over time, the goal line and a trend line which will be indicative of whether the team will complete all work by the end of the iteration.


For larger teams, especially those distributed, you need some way of tracking this in a centralised tool using TFS work items, FogBugz, VersionOne or a similar product which will allow the team to carry out work in the same way and automatically produce the burn down. In these circumstances it’s a good idea to have a project wall for those interested in progress – a large LCD on the wall rotating between different views on the iteration – burn down, work completed etc etc.

Ending the iteration
At the end of the iteration, the use cases should be complete, tested and essentially ready for delivery. At this point, the work can either be released through to acceptance testing or, for larger applications where it doesn’t add value to deliver small chunks, be demonstrated back to users. Testing is an interesting point – developers should be taking responsibility for their work and testing it to ensure all of the paths through the use case work as prescribed, a dedicated test team can augment this significantly and for larger developments it’s an absolute necessity. In this case, the output from the iteration should feed into the next testing iteration and defects should be fixed before it is released to user acceptance testing – the testing cycle (test, fix, test, uat, fat etc)  is a topic in itself and so I’ll leave it there for now.

In summary, engaging users and keeping them engaged is a critical aspect to the success of any development project, ignore users at your peril. I can testify that over numerous successful projects, keeping users engaged has been pivotal in successful delivery – the last thing you want to do is keep your project insular, deliver and then find that the finished product doesn’t match up to users expectations – that’s a sure fire way to ensure project failure.

Friday, 18 February 2011

Garbage collection in .NET – large and small object heap differences.

As promised in my previous post, I’ve just uploaded source code to codeplex (http://netgarbage.codeplex.com) which demonstrates the differences between the large and small object heaps and how allocating memory to the large object heap needs to be considered in applications that shifts large quantities of data around.

In this project you will find 4 main tests (not unit tests, I created this as a build and run console application);

  • TestInterlacedMemoryAllocationResultsInOutOfMemory
  • TestNonInterlacedMemoryAllocationDoesNotResultInOutOfMemory
  • TestInterlacdMemoryAllocationWithFullReleaseDoesNotResultInOutOfMemory
  • TestSmallObjectHeapInterlacedAllocationDoesCompaction


This test demonstrates one of the fundamental considerations of the large object heap. You will recall from my previous post that objects that are greater than 85000 bytes will be allocated onto the large object heap (LOH) instead of the small object heap (SOH) and the LOH does not compact memory when it collects unused memory. Rather it leaves behind holes. As such, allocating memory in a manner that is interleaved and then releasing every other block will leave small holes behind.

The net result of this is the application has lots of free memory overall, but it’s only available in small chunks. As allocations require the object being allocated to be laid out in memory in one contiguous block, trying to allocate an object that would appear to still fit in memory will fail as there is no hole in the LOH big enough to accommodate it. The result in this test is an out of memory exception;


The second test does exactly the same as the first, but rather than interlacing the memory it tries to allocate larger chunks together which means when one list of data is released and cleaned up, the allocation of large objects continues to function as there are holes in the LOH big enough to accommodate;


This test attempts to prove the reason that the first test fails with an OutOfMemory exception is because of the fragmentation – it does exactly the same as the first test, but releases both lists and sure enough, as expected, the further allocation succeeds.


For completeness I included this test to prove the difference between the SOH and the LOH. It does the same as the first test, allocates memory until it’s full by interleaving objects in two lists and then frees one list. (The objects are smaller to ensure they get allocated to SOH not LOH). As you will see, this test successfully allows memory to be allocated after just one list is freed because the garbage collector performs the compaction phase on the SOH, reshuffling memory to close the holes.


Get the source code here: http://netgarbage.codeplex.com

Thursday, 17 February 2011

Garbage collection fundamentals in .NET

Today I was asked a few questions about how garbage collection works, and it dawned on me that even though we’re on .NET 4.0 these days, there are a lot of .NET developers out there who don’t actually understand the mechanics of the garbage collector. So, even though this article is probably 10 years too late for a lot of people, I hope my rudimentary explanation will serve useful to those that don’t know about the concept or what it offers them. I’m going to cover;

  • What garbage collection is.
  • Some general information about the .Net garbage collector (GC)
  • What actually happens when you allocate an object.
  • The .NET garbage collection process
  • What the different generations are and what purpose they serve
  • How a GC gets triggered.

What is garbage collection?
Garbage collection provides an abstraction to the programmer that means they don’t need to worry about how memory is allocated and freed. In unmanaged code a programmer would need to allocate and free every byte their application used, using malloc() and free() (in C - new and delete in C++) and this has often been the source of numerous bugs in software, and they are often the hardest to track down.

Higher level languages like .NET take this burden away from the programmer by using garbage collection. The runtime takes care of tracking references to objects and understands when an object can be safely let go, so you can safely write;

void MyMethod()
    SomeLargeObject obj = new SomeLargeObject();

Instead of;

void MyMethod()
    SomeLargeObject obj = new SomeLargeObject;
    delete obj;

In the latter example, if you forgot to delete the allocated object, the memory it consumed would essentially be lost, a typical memory leak. With the former example, the garbage collection understand the object no longer has any active references and it gets garbage collected and disposed of during a GC cycle.

.NET garbage collection
In .NET the garbage collector runs in one of two modes – synchronous and concurrent (often referred to as workstation and server). When the GC determines it needs to perform some clean up, synchronous mode will pause the executing process, quickly clean up, and then resume the executable, whilst in concurrent mode (the default) the clean up executes in the background without blocking your application.

It’s important to note that every application has a maximum amount of memory that it can allocate objects to – this memory is known as the heap, and the size available is dependent upon the O/S. A 32 bit O/S can only allocate a maximum of 2Gb by default (can be increased to 3Gb with a boot config), whilst on a 64bit O/S, this is significantly higher (8Tb from memory, but don’t quote me).

The GC doesn’t run continually, and this is an important observation to make, it runs whenever certain conditions are met, which I’ll discuss shortly, but the important thing to note is when an object is no longer referenced it may not necessarily be collected immediately – rather it will be collected the next time the GC runs. This can be problematic for some objects that perhaps hold locks on contentious resources. Holding a file handle for instance – holding that for too long would be a bad thing as it could be a while before the handle is released freeing the file for access to other processes or other areas of your code.

To cope with this scenario .NET has the IDisposable interface. By implementing IDisposable, you write code to release these contentious resources when the object is no longer needed, but before the object holding them is actually collected. By wrapping the use of an IDisposable object within a using statement the dispose functionality will be called when the reference goes out of scope (when it drops out of the using block), thereby releasing the file handle in our example and making it available to other processes. You can also invoke the dispose functionality of an IDisposable manually by calling Dispose on the object, but the using statement is less prone to forgetfulness!

The actual memory used by the object will still only be released later, when the GC runs.

What happens when you allocate an object
Before we look further at the GC, it’s important to look at what happens when you “new up” an object. It is allocated on the managed heap, of which there are two with subtle differences – the small object heap and the large object heap.

The small object heap
Small objects ( those requiring less than 85,000 bytes) will be allocated on the small object heap. When the GC runs and reclaims space on the small object heap it does one thing different that the large object heap doesn’t – it compacts memory. It removes the holes left in the heap by the now vacant memory – more on this shortly.

The large object heap
When you allocate objects larger than 85,000 bytes (these are bytes within the object itself, not the objects it references), these are allocated on the large object heap, which doesn’t compact memory because of performance limitations (moving large chunks of memory around is slower), so it leaves holes in the heap and this should be considered in the design of your code, again I’ll explain why shortly.

Taking out the trash
When a garbage collection is triggered it runs through 2 phases – a marking phase and a compacting phase.

The marking phase is all about marking which objects are still in use and which ones aren’t. It starts by walking through all known objects from something called GC roots. A GC root is an anchor point on which objects are tracked – a GC root is any local variable in the currently running method, a static variable or managed objects passed to unmanaged code – these roots become the starting point for garbage collection, objects with finalizers are also roots and these take 2 GC cycles to get rid of – one to release another to finalize (although this can be changed by calling GC.SuppressFinalize from your dispose code).

So, the GC examines all objects to see if they can still be reached by any other live object, starting with the GC roots. If an object can be reached it is marked as still live and on it goes until it’s determined the state of all the objects being tracked. Each object is checked only once, if a reference has already been walked, it stops – this allows circular references to exist in your graphs.

At the end of the mark phase, the memory that wasn’t marked live is released and the compacting phase begins. This compacting phase only occurs on the small object heap – it doesn’t do this on the large object heap as it would be performance intensive (moving large chunks of memory around). The compacting phase is all about closing the holes in the heap. Take the diagram below - Let say you allocate 3 objects on the small object heap, named Obj1 to Obj3 – they are allocated sequentially onto the back of the heap. When Obj2 is no longer used, it gets cleaned up by the garbage collector and the compaction phase re-organises the heap to remove the hole that it left;


This has a number of advantages – first, when allocating a block of memory it must be allocated in one contiguous block otherwise the request to allocate will fail with an OutOfMemoryException. If you had holes in your heap of say 850 bytes each, but totalling 1,000,000 bytes free, you could still only allocate 850 bytes as the largest object because that would be the largest piece of contiguous memory available, so reshuffling the small object heap like this keeps your memory lean and mean.

A side effect of this is that new allocations become extremely quick as they are always to the end of the heap, there’s no need to scan for the space, the space is just there all in one block, there’s either enough space or there’s not.

The large object heap is a different beast. If you recall, any object over 85000 bytes will be allocated from the large object heap, which gets collected but doesn’t do the compaction. Taking our example above of 3 objects going on the large object heap you would have;


Notice the hole. If an allocation comes along that fits into this hole, it would be used, but no compaction is done. This can have implications – say you were to allocate two lists in an interleaved manner (so allocate a new object into list A, then a new one into list B) with each element on the list being 10mb. You do this say 100 times, so you allocate a total of 10 x 100 x 2 = 2000mb. Let’s say that’s the total amount of memory free and you’ve now stuffed your memory full to capacity.

Releasing one of these lists will give you back 1,000mb, but the largest chunk of memory available on the large object heap will be 10mb because the memory is in sequence and it doesn’t get compacted. As such, any allocation to the large object heap from that point that is over 10mb will fail with an OutOfMemoryException. So you see, when working with large amounts of data this is a consideration to bear in mind. I’ll explore this more in a future post with some sample code that I’ve written to demonstrate the point.

I should point out now that this is a slight oversimplification – the GC doesn’t actually check and mark every object, the GC runs against different generations.

What are these generations all about?
The GC in .NET has the concept of generations. A generation is a queue of the objects on the heap that represents, I’m tempted to say age of the object but that isn’t strictly true, rather it indicates the stickiness of an object. Freshly allocated objects on the small object heap are given generation 0 (there are three generations in total, 0, 1 and 2) indicating they are fresh and are therefore the most likely to be cleaned up next. Think about it, fresh objects you usually allocate, do something to them and then they are finished with.

So new objects, on the small object heap are given an initial generation of 0. When the GC runs, it can do so against a generation – generation 0 will clear only generation 0 objects that are no longer referenced, generation 1 will do 0 and 1 and finally generation 2 (known as a full garbage collection) will clear generations 0,1 and 2. As an object survives a garbage collection, it is promoted to the next generation up.

What this means is the next time a generation 0 collection is run, it won’t bother checking the promoted objects because it’s quite likely that those objects are still referenced, after all they survived one or more GC cycles already and so those objects won’t be collected until a more aggressive GC is executed. Promotions only occur of course up to generation 2, so if a gen 2 object lives through a GC cycle, it will still be gen 2.

The only additional thing to mention with this is that large objects (on the large object heap) are allocated as generation 2 directly whilst small objects are allocated as generation 0. You cannot manually allocate a generation 1 object.

When does a garbage collection start?
And so, the last thing to talk about is how a GC is initiated. There are 3 mechanisms;

  • Generation thresholds
  • Manual collection
  • Low memory conditions

Manual collection and low memory conditions
I’ll cover these two first as they are the simplest. If the hardware your code is running on is starting to run low on memory resources, your application will be notified and will automatically trigger a GC cycle – this is all handled for you, no code required, whilst the other mechanism is to manually call GC.Collect which will in turn perform a garbage collection cycle against the generation you passed as a parameter (plus any below it). This can be useful, but calling it is reasonably rare and you’re better off relying on the thresholds.

Each of the three generations has a GC threshold. When the heap is starting to fill with objects it may start to encroach on one of these thresholds and when it goes, the GC cycle for the generation in question will be triggered. This could result in object promotion to other generations which in turn may trigger further threshold collections. You don’t manage these thresholds directly, the runtime will dynamically tune the thresholds for you.

So that’s my hamfisted attempt at explaining the garbage collector in .NET, I hope it comes in handy for someone, it was certainly useful to refresh my own memory! I’ll post code in the next few days that demonstrates what I mentioned about the large object heap and fragmentation that will hopefully make some of this clearer.

Tuesday, 15 February 2011

SDLC – Productive development teams.

Since I started running and managing teams, about 10 years ago, I’ve been trying to put my finger on what it is that makes a productive development team. Of all the teams I’ve worked with, what makes one team perform better than the others if we consider that both teams are filled with equally capable members. Now, I’m not asking what makes the developers more productive as in what characteristics do they have personally, I’m asking what do we, as management, put in place in terms of environment and culture that allows one team to succeed more than another.

Over the years I’ve identified, read, and learned about a number of things that motivate technical teams, but I think there are two aspects that are the most crucial. We must provide technical teams with;

  • A quiet place to work with less interruption.
  • Authority and ownership of their work.

A quiet place to work
I’ve known this for 18 years, since I first became a developer - developers need a quiet place to work, with few interruptions. This fact is by far, THE most important thing we need to put in place to get the best out of a development team. These days I spend around half of my time doing non-development work; meetings, design sessions, writing specs, planning and all that kind of stuff, but the other half of my time I’ve usually got my head in visual studio, writing code, and I can tell you that a noisy working environment is still the primary thing that stops me from getting things done. All developers find it difficult to work when sat next to the call centre, or sat across from Jean from accounting who can’t stop telling everyone about how her two cats have wonderfully expressive expressions and how they smell divine when they’ve just been given a kitty bath – nice!

When a developer is writing code, they have in their head a plethora of details about what they are doing – class names, method names they need to call, variables they need to populate, and sequences they need to follow. They are thinking about the layer they are presently working in, the UI lets say, but are contemplating changes up the stack into repositories, domain objects and/or the database – all of this is being worked out, real time, in their heads.

Now, think about it, you’ve got this big jenga stack of information in your grey matter when Jean decides to tell you again about how Fuddles can smile when she says kitty kat – see what happens, your concentration is shattered, the jenga bricks all fall down and you now need to spend 5 or 10 minutes picking them up again – you essentially stall and get thrown out of  “The Zone” and it takes time to get back into it.

Joel Spolsky articulates this, and some of the math behind getting into and out of the Zone, in a much better way than I ever could, so I encourage you to read his articles or book  “Joel on software”.

So how can management address this quiet place to work issue? This is the tricky part – ideally you want each developer in their own office, with a door they can close when they want no interruptions, but this unfortunately usually results in laughter with the board who just don’t understand how a developer does their job. You can convey with explicit detail why but 9 times out of 10, they won’t put the partitions up.

At the very least, get your development team insulated from the rest of the business. One partition to create a “development office” will go a long way to reducing the noise distraction from the rest of the business in the open plan environment but this is only a 15% step in the right direction. You’re still going to have in-team distraction, where developer A decides it’s quicker to ask developer B a question rather than spend 2 minutes hunting down the email that gives him the answer and the partition will only be a minor deterrent from Dave from support waltzing in and interrupting everyone at the same time, but a 15% step in the right direction is still better than Joan and her cats.

Moral of the story – you might not be able to go the whole hog and get every developer into a private office, but do something.

Authority and ownership
In my current role, I’m assigned a project to take on, I get a briefing and then I’m left to it. Sure there are governance reviews every few months, we usually have a project manager assigned to interface the client commercially and do the paper pushing such as weekly status reports back to the businesses (ours and the clients) to keep everyone happy, but from a technical design, development and team management perspective, I’m on my own – I take full ownership and responsibility for everything we do, and that my friend is empowerment which breeds motivation!

Now consider Kevin. Kevin works for the antithesis of my company and they are total control freaks, they need to micro manage the soul out of Kevin so that they know he’s doing exactly what they want him to do – even if they are the one’s in the worst position to be making technical decisions! Kevin gets on with code, line by line, writing to specification exactly what the management team are looking for, they check up on him hourly, tell him to change his code to work slightly differently and eventually Kevin either quits or comes into work with an AK47 and mows down the management team.

That’s an extreme example, but I have honestly seen that – well, not the AK47 part. I’ve been in more than one organisation before where the developers and architects are micro managed from up on high, told which technologies to use because the On-High-Manager had a conversation with someone the day before and decided that Service-Orientated-Entity-Hibernate was the up and coming technology that is the silver bullet for all their problems – the development team protest but are shot down and told to just get on with it and what do you know, 2 months later, they call in someone like me to help them work out what their problems are.

Developers need to own their work. Sure, architects and SDMs can put in place frameworks and prescribed approaches, but ultimately developers are a creative breed and so let them be creative, don’t confine them to a hell of micro management, give them a task, set some boundaries (the general architecture) and let them get on with it.

Now, I can hear some of those micro managers screaming now “what about the slackers” and the “people that take the proverbial”. Well, there’s three things I’d say to that. Firstly your recruitment process is wrong, Second, the reason they are slacking is more likely because of their lack of motivation due to being brow beaten daily and Three, your lightweight management process will weed out those who just, bluntly, aren’t up to standard.

Recruitment process
I say your recruitment process is wrong because if you have slackers or proverbial takers on your team, your process wasn’t good enough to weed them out. Did a couple of technical and highly motivated, passionate people interview them? Were they given a technical test to test their programming skills? Were they given a logical test to test their thought and reasoning? If not, your recruitment process is wrong and you’re going to let through some people who should be doing something else with their time, but shouldn’t be writing code for a living.

Brow beaten
When you’re running your development team in micro management mode, developers, who are creative by nature, start to lose motivation. When they have a spec that says, write some code, like this, that is so detailed that they have no creative input, they would probably sooner spend a day on facebook than actually write that stinking code you’ve told them to write. This is not the fault of the developer, this is poor management – it’s your job to remove things from their path that stop them doing their job and motivate them to do their best work whilst their job is to write beautiful code.

Lightweight management process to weed out the useless
So whilst I’m saying have more of a hands-off approach to management, you will still have a management method that will allow you to compare apples with apples. SCRUM is obviously one of my favourite approaches to running iterative development and it’s this that I think gives you a good tool to be able to monitor the performance of all developers and identify those who are struggling.

By working out what work was delivered by each individual in each sprint, you get not only a team velocity (how much work a team can generally carry out in each iteration), but also a velocity per team member. Now if Dave and John are hitting 40 points in a 2 week sprint but Alan is only hitting 10, you’ve got to ask why Alan’s productivity is so low – you don’t need to be all “up’in’Alan’s’Bizniss'” and micro manage him, you just need to identify his velocity slump and then ask what the problem is – if it’s because his cat died this week and so couldn’t get in the Zone, then fine, send him to grief counselling, but if it’s because he just can’t do the job, give him some opportunities to skill up but then I’m afraid it’s time to weed him out – you only want people on your team who get things done.

I’m not saying these are the only two things that you need to do to motivate a development team, but I am saying they are the two most important out of a long list. There’s more to the list, but others smarter than I have already covered this in detail (comfy chairs, large monitors, powerful pc’s, interesting work, the best tools and so on). If you can do only two things to motivate your team, I strongly urge you to do the two things above.

Monday, 14 February 2011

Prism, MVVM and Silverlight – the finished article.

In my previous article I set out to create a simple Silverlight application using the prism framework and the MVVM pattern. I chose what I thought would be quite a straight forward application to build, something that would mirror the iPad application, Popplet (links: Website, get it on iTunes) – which as I said in my earlier post, is a pretty neat little app that lets you essentially drop post it notes (text or image) onto a canvas, drag them around into some layout that makes sense to you and connect the various boxes.

iPad Screenshot 1

So, this distilled down to the following user stories;

  • As a user I want to see a canvas of notes.
  • As a user I want to be able to add a new note box to my canvas.
  • As a user I want to be able to rotate or resize any of my note boxes on my canvas
  • As a user I want to be able to connect note boxes
  • As a user I want to be able to delete note boxes
  • As a user I want to be able to zoom in and out
  • As a user I want to be able to drag my canvas around
  • As a user I want to be able to click a button to zoom out to the extents of my canvas so far.

I’ve spent about 12/13 hours building this in total, over 5 sessions, from scratch, learning some of the concepts as I’ve gone and I’ve delivered all of the user stories above with the exception of the ability to connect note boxes - basically I’ve run out of time to add that, but it should be trivial to add by creating another canvas layer in our control template that is used to render connector lines.

This was an exercise not only in how best to work with Prism and MVVM but also in how I might best implement a reasonably complex control – the designer surface and whilst I’m impressed with how quickly I was able to build this little app (I think it could be done within a day easily), I also think that my approach is actually the wrong one.

The way this application works is we have a custom control, the diagram canvas which contains all the functionality for zooming, panning and so on, and then the items themselves are rendered using a data template. Two data templates in fact – one that dictates how the standard node should be rendered, the other dictating the rendering of it when it’s selected – in which case it is adorned with resize handles, rotation handles, a button to delete the entity and a textbox to change the text (there is no adornment layer in SilverLight it turns out!).

This all works, but it doesn’t feel very flexible. It’s pretty much impossible to represent more than one shape type – the list of “stuff” you bind to all gets rendered through one data template, so that makes life difficult.

It also feels quite tightly coupled and doesn’t have clean separation of concerns (eg the diagram component has code in there to find UI elements within a data template that should be refactored out). Anyway, the exercise has proven it’s point, but if I were to do this again, I’d try to be a little more flexible and use different controls (+control templates) for different types of shape (all of these shape controls would implement IShape) and expose a property that describes their adornments – so some IShapes would adorn themselves with a RotatorAdornment and SizerAdornment whilst others might just have EditTextAdornment and so on. Composing the controls in this way would be much more flexible, and the view model would have the responsibility of transforming our domain data (the model) into the shape data collection the canvas is binding to and vice versa.

Anyway, for now, this is what the application looks like, it works, it does what it says on the tin, but it’s perhaps not the best implementation. You can grab the source from http://silverpop.codeplex.com.


Tuesday, 8 February 2011

Hmmm– what happened to my brain?

I’ve been reading the excellent “Joel on software” and “More Joel on software” books recently (by Joel Spolsky obviously) – I’ve got to say I wish more people had his insight into what motivates and drives a good developer to work well. I’ve lost count of the number of clients I’ve turned up at to find brow beaten developers sat next to the telesales department staring at postage stamp screens and wondering why their productivity is so low….

Anyway, I’ve got a few things to say about that myself, but I’ll save it for another day, today I’m thinking about some of the stuff Joel talks about with regards to groking pointers and how programming is getting easier, or should I say dumbed down (due to University teaching higher languages such as Java rather than C, C++ or assembly)  and I looked back to where I was in the early ‘90s and realised that I seem to have forgotten a lot since then!

Back then, I remember writing games in my spare time in C/C++ and assembler – I remember knowing about memory segments and offsets, I remember using registers to set source/target memory and moving bytes around, I “got” pointers, I knew about memory allocation, doubly linked lists, and I could write assembly language until it came out of my ears - I even remember using “debug” on the command line to inspect what a game was doing during the license key check at the start up of the game and then changing the op codes from a JE (jump if equal) to a JNE (jump if not equal) so that if you put the wrong key in, it let you play the game (sorry about that, but we’d erm, lost the key, that’s it, lost it).

I even remember myself and a few colleagues (John Steele and David Mort, just in case you’re out there!) having competitions on writing double buffer, screen swapping algorithms that would push the delta from two buffers into VGA video memory (0xA000)– because writing to video memory was so unbearably slow compared to RAM – we would actually sit there, pouring over the assembly listings, counting the iterations and how many clock cycles it would take to refresh a partial or full screen and therefore who’s routine was better. Geek fest, and you know what, I loved it and feel sorry for any developers out there that haven’t had a similar experience.

Anyway - what is my point? Well, I don’t think I have one other than this was a nice trip down memory lane, and I think I might struggle to go back and do that again now. I could do it with my eyes closed back in ‘92 and I had to for a variety of, usually performance, reasons, but because I’ve not had to use that skill in a long time, it’s gotten old and rusty like a neglected banger of a car left in a leaky garage. These days I work at a higher level of abstraction – I work with C# and I build software in that higher level world rather than in the bowels with assembly and whilst there are numerous principles and concepts learned in the past that are still useful, I’ve not needed to drop into assembly for any reason (building LOB apps) in the last, I’m guessing, 10 years.

Is that a good thing? Part of me is happy that building software is easier, but the other part of me feels that those kinds of skills provide an essential grounding to really understanding what is going on under the hood, even if you do spend your days in the .NET C# clouds.

Friday, 4 February 2011

SDLC – agile in the real world – consulting and long complex projects (Iteration 0)

Building software isn’t like building a bridge. This is not new thinking, this is well understood and has been a factor of software development since the 70’s.

When building a bridge someone has a vision that is part creative and part engineering on how the bridge should be built. The engineers then take this vision and run calculations, build models, run stress tests and so on to make sure that the design will stand up in the real world. They then produce a detailed schedule of work on what needs to happen each week to get the bridge built – there is no deviation from the plan, no changing your mind about how the bridge should be implemented and, beyond the initial vision and problem solving – very little creativity.

Building software on the other hand is a creative process. Developers who write code can approach a problem in a number of different ways, some good, some bad, some passable – but all creative, drawn from their bag of experience and knowledge.  Yes, it’s engineering, but I propose it’s also extremely creative.

The case for up-front-design (but not BIG up-front-design)
I’ll even go so far as to argue that planning a bridge is easier than planning software – we already know every little detail about the bridge design up front, we know that design isn’t going to change and we know it will take 4 hours for Dave with a spanner to attach two I beams together, typically it takes 3 hours to weld the plates that make a section, but sometimes that weld is poor, so you build in a contingency factor and so on – easy. (Aside: for any bridge builders out there who just happen to have stumbled on this post, I’m sure it’s really not that easy!)

Planning and managing software development on the other hand – well, I’m an advocate of SCRUM and agile approaches, sure, and that’s a great way to run a project,  but I’m not so sure that it’s a great way to start a project from the beginning. XP would have us diving straight into code, SCRUM we’d already have a backlog to pick some items off ready to start building in iteration 1, so I guess this post is about how that backlog get’s filled to start with – an iteration 0 - just diving headlong into code without first understanding the holistic picture of the app you’re going to build, the pieces that will make up the whole and the general stages you need to go through is a recipe for disaster on a project that lasts say 2 years with 3 workstreams (custom dev, business intelligence and data migration).

Also, I’m a consultant and I generally work with a team of consultants from my company (often with other developers supplied by the client or with an offshore development team) and these people cost money – it’s extremely rare for a client to engage say 4 consultants at £1,000 a day without having some sort of idea of how long we’ll take to deliver and how much it will cost.

So, I’m sorry XP purists, but you can’t escape the fact that we need up front design.

Up-front-design in perspective
I’m not saying we need big thick, 000’s page long manuals with detailed conceptual, logical and physical designs down to sequence diagrams for individual methods, no, but we do need to;

  1. Gain insight into the overall application surface
  2. Identify the pieces of the application – essentially the sub-applications within the whole – we’ll call these domain components.
  3. Flesh the high level use cases/user stories for each domain component.

From this the project can be planned into iterations and roughly estimated – the estimate won’t be 100% but it will be an indication and a yardstick to measure change by. In other words, if the customer during a sprint decides to change tack on part of the application, its quite easy to see that adding that extra work to the next iteration will make something fall off the back end of the project and either not get delivered or will cost more. The exercise of defining the high level project will give you a good idea of how best to structure delivery in iterations and let you populate a backlog so you can see the effect of change and re-plan accordingly during the execution of the build.

Alongside the identification of the big picture, the application architect will also be considering the approach to development – will the user want a rich UI interface, will it be a desktop application, a web based application, maybe SilverLight would be a good idea – should the business logic be distributed and exposed as services – will this application expose itself to other applications – how will we handle workflow – should we surface through a SharePoint portal? - all these things will become clearer as the high level design above is being done, but at the same time, the architect will be making some technical decisions that are based on assumption and are therefore risky for the success of the project – we need a way to mitigate that risk, less we head off in the wrong direction only to find that the assumption was wrong in 10 months and have a lot of re-engineering to do to fix it!

Projects need iterations but they also need phases
All of this indicates what we already know - that projects need phases. They need this start up phase to identify the big picture, estimate the whole and de-risk technical assumptions, but what are the phases of the project from there? Well, if you subscribe to any of the methodologies like RUP, Open/UP, MSF etc, you will already be familiar with project phases – the list below describes the phases of a project how I personally like to work (but I’m not precious or religious about these in any way – call them whatever works for you!).

  1. Inception
    • During this phase you identify the big picture – application surface, domain components and drill into use case titles/user stories
    • You also identify the proposed architecture for the solution
    • Build a proof of concept to start to de-risk some of the technical assumptions
    • Deliver
      • 30,000ft map of the entire application
      • High level use cases
      • Iteration plan
      • Proposed costs and timescales
      • Use case detail, wireframes and specification detail for the first elaboration sprint.
  2. Elaboration
    • This phase we start to work in iterations, using SCRUM to manage those iterations.
      • Each iteration delivers the use case detail (building on the high level headers), lo-fidelity screen wireframes and other pertinent specification detail for the next iteration.
    • De-risk any additional technical risks
    • Build a fully functioning vertical slice through the architecture and problem domain
  3. Construction
    • SCRUM sprints/iterations through the build of the product.
    • Stringent quality control – continuous builds, nightly build released to testing.
    • Continue to provide design detail 1 sprint ahead.
  4. Transition
    • Releasing the product – a release sprint.
      • In some cases you can’t release at the end of each 2-4 week sprint – the application may be too large and what you can build in an iteration too small.
      • Instead each iteration should deliver demoware – a presentation or demo to some of the users.
    • Some applications can be extremely complex involving multiple deployment steps, these should be planned.
  5. Support
    • Handing over to the people responsible for the upkeep of the system.

In future posts I’ll cover some of these phases, how we go about running the inception phase for example, key things to do in workshops with customers, and how best to start the process of identifying the big picture.

To conclude though, I will say this - yes, the above is applicable to larger sized projects. If you’re writing a simple expense tracking application for the sales department or a personal journal application for yourself then yes, go pure agile and start coding. If you’re working in a larger team on a larger project, you’re going to have to put some effort in up front I’m afraid to get the project rolling – you can run the project as SCRUM from then on and get all the juicy benefits from it, whilst also understanding the bigger picture, budget and timescale implications.