Specific down-sides to many-'small'-assemblies?

Question

I am planning out some work to introduce Dependency Injection into what is currently a large monolithic library in an attempt to make the library easier to unit-test, easier to understand, and possibly more flexible as a bonus.

I have decided to use NInject, and I really like Nate's motto of 'do one thing, do it well' (paraphrased), and it seems to go particularly well within the context of DI.

What I have been wondering now, is whether I should split what is currently a single large assembly into multiple smaller assemblies with disjoint feature sets. Some of these smaller assemblies will have inter-dependencies, but far from all of them, because the architecture of the code is pretty loosely coupled already.

Note that these feature sets are not trivial and small unto themselves either... it encompasses things like client/server communications, serialisation, custom collection types, file-IO abstractions, common routine libraries, threading libraries, standard logging, etc.

I see that a previous question: What is better, many small assemblies, or one big assembly? kind-of addresses this issue, but with what seems to be even finer granularity that this, which makes me wonder if the answers there still apply in this case?

Also, in the various questions that skirt close to this topic a common answer is that having 'too many' assemblies has caused unspecified 'pain' and 'problems'. I would really like to know concretely what the possible down-sides of this approach could be.

I agree that adding 8 assemblies when before only 1 was needed is 'a bit of a pain', but having to include a big monolithic library for every application is also not exactly ideal... plus adding the 8 assemblies is something you do only once, so I have very little sympathy for that argument (even tho I would probably complain along with everyone else at first).

Addendum:
So far I have seen no convinging reasons against smaller assemblies, so I think I will proceed for now as if this is a non-issue. If anyone can think of good solid reasons with verifiable facts to back them up I would still be very interested to hear about them. (I'll add a bounty as soon as I can to increase visibility)

EDIT: Moved the performance analysis and results into a separate answer (see below).

Answer 1

Since the performance analysis has become a little lengthier than expected, I've put it into it's own separate answer. I will be accepting Peter's answer as official, even though it lacked measurements since it was most instrumental in motivating me to perform the measurements myself, and since it gave me the most inspiration for what might be worth measuring.

Analysis:
The concrete downsides mentioned so far seem to all focus on performance of one kind of another but actual quantitative data was missing, I have done some measurements of the following:

• Time to load solution in the IDE
• Time to compile in the IDE
• Assembly load time (time it takes the application to load)
• Lost code optimisations (time it takes an algorithm to run)

This analysis completely ignores the 'quality of the design', which some people have mentioned in their answers, since I do not consider the quality a variable in this trade-off. I am assuming that the developer will first and foremost let their implementation be guided by the desire to get the best possible design. The trade-off here is whether it is worthwhile aggregating functionality into larger assemblies than the design strictly calls for, for the sake of (some measure of) performance.

Application structure:
The application I built is somewhat abstract because I needed a large number of solutions and projects to test with, so I wrote some code to generate them all for me.

The application contains 1000 classes, grouped into 200 sets of 5 classes that inherit from each other. Classes are named Axxx, Bxxx, Cxxx, Dxxx and Exxx. Classes A is completely abstract, B-D are partially abstract, overriding one of the methods of A each, and E is concrete. The methods are implemented so that a call of one method on instances of E will perform multiple calls up the hierarchy chain. All method bodies are simple enough that they should theoretically all inline.

These classes were distributed across assemblies in 8 different configurations along 2 dimensions:

• Number of assemblies: 10, 20, 50, 100
• Cutting direction: across the inheritance hierarchy (none of A-E are ever in the same assembly together), and along the inheritance hierarchy

The measurements are not all exactly measured; some were done by stopwatch and have a larger margin of error. The measurements taken are:

• Opening the solution in VS2008 (stopwatch)
• Compiling the solution (stopwatch)
• In IDE: Time between start and first executed line of code (stopwatch)
• In IDE: Time to instantiate one of Exxx for each of the 200 groups in the IDE (in code)
• In IDE: Time to execute 100,000 invocations on each Exxx in the IDE (in code)
• The last three 'In IDE' measurements, but from the prompt using the 'Release' build

Results:
Opening the solution in VS2008

                               ----- in the IDE ------   ----- from prompt -----
Cut    Asm#   Open   Compile   Start   new()   Execute   Start   new()   Execute
Across   10    ~1s     ~2-3s       -   0.150    17.022       -   0.139    13.909
         20    ~1s       ~6s       -   0.152    17.753       -   0.132    13.997
         50    ~3s       15s   ~0.3s   0.153    17.119    0.2s   0.131    14.481
        100    ~6s       37s   ~0.5s   0.150    18.041    0.3s   0.132    14.478

Along    10    ~1s     ~2-3s       -   0.155    17.967       -   0.067    13.297
         20    ~1s       ~4s       -   0.145    17.318       -   0.065    13.268
         50    ~3s       12s   ~0.2s   0.146    17.888    0.2s   0.067    13.391
        100    ~6s       29s   ~0.5s   0.149    17.990    0.3s   0.067    13.415

Observations:

• The number of assemblies (but not the cutting direction) seems to have a roughly linear impact on the time it takes to open the solution. This does not strike me as surprising.
• At about 6 seconds, the time it takes to open the solution does not seem to me an argument to limit the number of assemblies. (I did not measure whether associating source control had a major impact on this time).
• Compile time increases a little more than linearly in this measurement. I imagine most of this is due to the per-assembly overhead of compilation, and not inter-assembly symbol resolutions. I would expect less trivial assemblies to scale better along this axis. Even so, I personally don't find 30s of compile time an argument against splitting, especially when noting that most of the time only some assemblies will need re-compilation.
• There appears to be a barely measurable, but noticeable increase in start-up time. The first thing the application does is output a line to the console, the 'Start' time is how long this line took to appear from start of execution (note these are estimates because it was too quick to measure accurately even in worst-case).
• Interestingly, it appears that outside the IDE assembly loading is (very slightly) more efficient than inside the IDE. This probably has something to do with the effort of attaching the debugger, or some such.
• Also note that re-start of the application outside the IDE reduced the start-up time a little further still in the worst-case. There may be scenarios where 0.3s for start-up is unacceptable, but I cannot imagine this will matter in many places.
• Initialisation and execution time inside the IDE are solid regardless of the assembly split-up; this may be a case of the fact that it needs to debug causing it to have an easier time at resolving symbols across assemblies.
• Outside the IDE, this stability continues, with one caveat... the number of assemblies does not matter for the execution, but when cutting across the inheritance hierarchy, the execution time is a fraction worse than when cutting along. Note that the difference appears too small to me to be systemic; it probably is extra time it takes the run-time once to figure out how to do the same optimisations... frankly although I could investigate this further, the differences are so small that I am not inclined to worry too much.

So, from all this it appears that the burden of more assemblies is predominantly borne by the developer, and then mostly in the form of compilation time. As I already stated, these projects were so simple that each took far less than a second to compile causing the per-assembly compilation overhead to dominate. I would imagine that sub-second assembly compilation across a large number of assemblies is a strong indication that these assemblies have been split further than is reasonable. Also, when using pre-compiled assemblies the major developer argument against splitting (compilation time) would also disappear.

In these measurements I can see very little if any evidence against splitting into smaller assemblies for the sake of run-time performance. The only thing to watch out for (to some extent) is to avoid cutting across inheritance whenever possible; I would imagine that most sane designs would limit this anyway because inheritance would typically only occur within a functional area, which would normally end up within a single assembly.

Answer 2

I will give you a real-world example where the use of many (very) small assemblies has produced .Net DLL Hell.

At work we have a large homegrown framework that is long in the tooth (.Net 1.1). Aside from usual framework type plumbing code (including logging, workflow, queuing, etc), there were also various encapsulated database access entities, typed datasets and some other business logic code. I wasn't around for the initial development and subsequent maintenance of this framework, but did inherit it's use. As I mentioned, this entire framework resulted in numerous small DLLs. And, when I say numerous, we're talking upwards of 100 -- not the managable 8 or so you've mentioned. Further complicating matters were that the assemblies were all stronly-signed, versioned and to appear in the GAC.

So, fast-forward a few years and a number of maintenance cycles later, and what's happened is that the inter dependencies on the DLLs and the applications they support has wreaked havoc. On every production machine is a huge assembly redirect section in the machine.config file that ensures that "correct" assembly get's loaded by Fusion no matter what assembly is requested. This grew out of the difficulty that was encountered to rebuild every dependent framework and application assembly that took a dependency on one that was modified or upgraded. Great pains (usually) were taken to ensure that no breaking changes were made to assemblies when they were modified. The assemblies were rebuilt and a new or updated entry was made in the machine.config.

Here's were I will pause to listen to the sound of a huge collective groan and gasp!

This particular scenario is the poster-child for what not to do. Indeed in this situation, you get into a completely unmaintainable situation. I recall it took me 2 days to get my machine setup for development against this framework when I first started working with it -- resolving differences between my GAC and a runtime environment's GAC, machine.config assembly redirects, version conflicts at compile time due to incorrect references or, more likely, version conflict due to direct referencing component A and component B, but component B referenced component A, but a different version than my application's direct reference. You get the idea.

The real problem with this specific scenario is that the assembly contents were far too granular. And, this is ultimately what caused the tangled web of inter dependencies. My thoughts are that the initial architects thought this would create a system of highly maintainable code -- only having to rebuild very small changes to components of the system. In fact, the opposite was true. Further, to some of the other answers posted here already, when you get to this number of assemblies, loading a ton of assemblies does incur a performance hit -- definitely during resolution, and I would guess, though I have no empirical evidence, that runtime might suffer in some edge case situations, particularly where reflection might come into play -- could be wrong on that point.

You'd think I'd be scorned, but I believe there are logic physical separations for assemblies -- and when I say "assemblies" here, I am assuming one assembly per DLL. What it all boils down to are the inter dependencies. If I have an assembly A that depends on assembly B, I always ask myself if I'll ever have the need to reference assembly B with out assembly A. Or, is there a benefit to that separation. Looking at how assemblies are referenced is usually a good indicator as well. If you were to divide your large library in assemblies A, B, C, D and E. If you referenced assembly A 90% of the time and because of that, you always had to reference assembly B and C because A was dependent on them, then it's likely a better idea that assemblies A, B and C be combined, unless there's a really compelling argument to allow them to remain separated. Enterprise Library is classic example of this where you've nearly always got to reference 3 assemblies in order to use a single facet of the library -- in the case of Enterprise Library, however, the ability to build on top of core functionality and code reuse are the reason for it's architecture.

Looking at architecture is another good guideline. If you have a nice cleanly stacked architecture, where your assembly dependencies are int the form of a stack, say "vertical", as opposed to a "web", which starts to form when you have dependencies in every direction, then separation of assemblies on functional boundaries makes sense. Otherwise, look to roll things into one or look to re-architect.

Either way, good luck!

Answer 3

There is a slight performance hit to loading each assembly (even more if they are signed), so that's one reason to tend to cluster commonly-used things together in the same assembly. I don't believe there's a big overhead once things are loaded (though there may be some static optimization stuff that the JIT may have a harder time performing when crossing an assembly boundary).

The approach I try to take is this: Namespaces are for the logical organization. Assemblies are to group classes/namespaces that should be physically used together. Ie. if you don't expect to want ClassA and not ClassB (or vice versa), they belong in the same assembly.