post

UPDATE: SCons is Still Really Slow

A while back I posted a series of articles exploring the scalability of SCons, a popular Python-based build tool. In a nutshell, my experiments showed that SCons exhibits roughly quadratic growth in build runtimes as the number of targets increases:

Recently Dirk Baechle attempted to rebut my findings in an entry on the SCons wiki: Why SCons is not slow. I thought Dirk made some credible suggestions that could explain my results, and he did some smart things in his effort to invalidate my results. Unfortunately, his methods were flawed and his conclusions are invalid. My original results still stand: SCons really is slow. In the sections that follow I’ll share my own updated benchmarks and show where Dirk’s analysis went wrong.

Test setup

As before, I used genscons.pl to generate sample builds ranging from 2,000 to 50,000 targets. However, my test system was much beefier this time:

2013 2010
OS Linux Mint 14 (kernel version 3.5.0-17-generic) RedHat Desktop 3 (kernel version 2.4.21-58.ELsmp)
CPU Quad 1.7GHz Intel Core i7, hyperthreaded Dual 2.4GHz Intel Xeon, hyperthreaded
RAM 16 GB 2 GB
HD SSD (unknown)
SCons 2.3.0 1.2.0.r3842
Python 2.7.3 (system default) 2.6.2

Before running the tests, I rebooted the system to ensure there were no rogue processes consuming memory or CPU. I also forced the CPU cores into “performance” mode to ensure that they ran at their full 1.7GHz speed, rather than at the lower 933MHz they switch to when idle.

Revisiting the original benchmark

I think Dirk had two credible theories to explain the results I obtained in my original tests. First, Dirk wondered if those results may have been the result of virtual memory swapping — my original test system had relatively little RAM, and SCons itself uses a lot of memory. It’s plausible that physical memory was exhausted, forcing the OS to swap memory to disk. As Dirk said, “this would explain the increase of build times” — you bet it would! I don’t remember seeing any indication of memory swapping when I ran these tests originally, but to be honest it was nearly 4 years ago and perhaps my memory is not reliable. To eliminate this possibility, I ran the tests on a system with 16 GB RAM this time. During the tests I ran vmstat 5, which collects memory and swap usage information at five second intervals, and captured the result in a log.

Next, he suggested that I skewed the results by directing SCons to inherit the ambient environment, rather than using SCons’ default “sanitized” environment. That is, he felt I should have used env = Environment() rather than env = Environment(ENV = os.environ). To ensure that this was not a factor, I modified the tests so that they did not inherit the environment. At the same time, I substituted echo for the compiler and other commands, in order to make the tests faster. Besides, I’m not interested in benchmarking the compiler — just SCons! Here’s what my Environment declaration looks like now:

env = Environment(CC = 'echo', AR = 'echo', RANLIB = 'echo')

With these changes in place I reran my benchmarks. As expected, there was no change in the outcome. There is no doubt: SCons does not scale linearly. Instead the growth is polynomial, following an n1.85 curve. And thanks to the the vmstat output we can be certain that there was absolutely no swapping affecting the benchmarks. Here’s a graph of the results, including an n1.85 curve for comparison — notice that you can barely see that curve because it matches the observed data so well!

SCons full build runtime - click for larger view

For comparison, I used the SCons build log to make a shell script that executes the same series of echo commands. At 50,000 targets, the shell script ran in 1.097s. You read that right: 1.097s. Granted, the shell script doesn’t do stuff like up-to-date checks, etc., but still — of the 3,759s average SCons runtime, 3,758s — 99.97% — is SCons overhead.

I also created a non-recursive Makefile that “builds” the same targets with the same echo commands. This is a more realistic comparison to SCons — after all, nobody would dream of actually controlling a build with a straight-line shell script, but lots of people would use GNU make to do it. With 50,000 targets, GNU make ran for 82.469s — more than 45 times faster than SCons.

What is linear scaling?

If the performance problems are so obvious, why did Dirk fail to see them? Here’s a graph made from his test results:

SCons full build runtime, via D. Baechle - click for full size

Dirk says that this demonstrates “SCons’ linear scaling”. I find this statement baffling, because his data clearly shows that SCons does not scale linearly. It’s simple, really: linear scaling just means that the build time increases by the same amount for each new target you add, regardless of how many targets you already have. Put another way, it means that the difference in build time between 1,000 targets and 2,000 targets is exactly the same as the difference between 10,000 and 11,000 targets, or between 30,000 and 31,000 targets. Or, put yet another way, it means that when you plot the build time versus the number of targets, you should get a straight line with no change in slope at any point. Now you tell me: does that describe Dirk’s graph?

Here’s another version of that graph, this time augmented with a couple additional lines that show what the plot would look like if SCons were truly scaling linearly. The first projection is based on the original graph from 2,500 to 4,500 targets — that is, if we assume that SCons scales linearly and that the increase in build time between 2,500 and 4,500 targets is representative of the cost to add 2,000 more targets, then this line shows us how we should expect the build time to increase. Similarly, the second projection is based on the original graph between 4,500 and 8,500 targets. You can easily see that the actual data does not match either projection. Furthermore you can see that the slope of these projections is increasing:

SCons full build runtime with linear projections, via D. Baechle - click for full size

This shows the importance of testing at large scale when you’re trying to characterize the scalability of a system from empirical data. It can be difficult to differentiate polynomial from logarithmic or linear at low scales, especially once you incorporate the constant factors — polynomial algorithms can sometimes even give better absolute performance for small inputs than linear algorithms! It’s not until you plot enough data points at large enough values, as I’ve done, that it becomes easy to see and identify the curve.

What does profiling tell us?

Next, Dirk reran some of his tests under a profiler, on the very reasonable assumption that if there was a performance problem to be found, it would manifest in the profiling data — surely at least one function would demonstrate a larger-than-expected growth in runtime. Dirk only shared profiling data for two runs, both incremental builds, at 8,500 and 16,500 targets. That’s unfortunate for a couple reasons. First, the performance problem is less apparent on incremental builds than on full builds. Second, with only two datapoints it is literally not possible to determine whether growth is linear or polynomial. The results of Dirk’s profiling was negative: he found no “significant difference or increase” in any function.

Fortunately it’s easy to run this experiment myself. Dirk used cProfile, which is built-in to Python. To profile a Python script you can inject cProfile from the command-line, like this: python -m cProfile scons. Just before Python exits, cProfile dumps timing data for every function invoked during the run. I ran several full builds with the profiler enabled, from 2,000 to 20,000 targets. Then I sorted the profiling data by function internal time (time spent in the function exclusively, not in its descendents). In every run, the same two functions appeared at the top of the list: posix.waitpid and posix.fork. To be honest this was a surprise to me — previously I believed the problem was in SCons’ Taskmaster implementation. But I can’t really argue with the data. It makes sense that SCons would spend most of its time running and waiting for child processes to execute, and even that the amount of time spent in these functions would increase as the number of child processes increases. But look at the growth in runtimes in these two functions:

SCons full build function time, top two functions - click for full size

Like the overall build time, these curves are obviously non-linear. Armed with this knowledge, I went back to Dirk’s profiling data. To my surprise, posix.waitpid and posix.fork don’t even appear in Dirk’s data. On closer inspection, his data seems to include only a subset of all functions — about 600 functions, whereas my profiling data contains more than 1,500. I cannot explain this — perhaps Dirk filtered the results to exclude functions that are part of the Python library, assuming that the problem must be in SCons’ own code rather than in the library on which it is built.

This demonstrates a second fundamental principle of performance analysis: make sure that you consider all the data. Programmers’ intuition about performance problems is notoriously bad — even mine! — which is why it’s important to measure before acting. But measuring won’t help if you’re missing critical data or if you discard part of the data before doing any analysis.

Conclusions

On the surface, performance analysis seems like it should be simple: start a timer, run some code, stop the timer. Done correctly, performance analysis can illuminate the dark corners of your application’s performance. Done incorrectly — and there are many ways to do it incorrectly — it can lead you on a wild goose chase and cause you to squander resources fixing the wrong problems.

Dirk Baechle had good intentions when he set out to analyze SCons performance, but he made some mistakes in his process that led him to an erroneous conclusion. First, he didn’t run enough large-scale tests to really see the performance problem. Second, he filtered his experimental data in a way that obscured the existence of the problem. But perhaps his worst mistake was to start with a conclusion — that there is no performance problem — and then look for data to support it, rather than starting with the data and letting it impartially guide him to an evidence-based conclusion.

To me the evidence seems indisputable: SCons exhibits roughly quadratic growth in runtimes as the number of build targets increases, rendering it unusable for large-scale software development (tens of thousands of build outputs). There is no evidence that this is a result of virtual memory swapping. Profiling suggests a possible pair of culprits in posix.waitpid and posix.fork. I leave it to Dirk and the SCons team to investigate further; in the meantime, you can find my test harness and test results in my GitHub repo. If you can see a flaw in my methodology, sound off in the comments!

post

Why is SCons so slow?

UPDATE: If you’re coming from Why SCons is not slow, you should read my response

A while back, I did a series of posts exploring the performance of SCons on builds of various sizes. The results were dismal: SCons demonstrated a classic O(n2) growth in runtime, meaning that the length of the build grew in proportion to the square of the number of files in the build, rather than linearly as one would hope. Naturally, that investigation and its results provoked a great deal of discussion at the time and since. Typically, SCons advocates fall back on one particular argument: “Sure, SCons may be slow,” they say, “but that’s the price you pay for a correct build.” Recently, Eric S. Raymond wrote an article espousing this same fundamental argument, with the addition of some algorithmic analysis intended to prove mathematically that a correct build, regardless of the build tool, must necessarily exhibit O(n2) behavior — a clever bit of circular logic, because it implies that any build tool that does not have such abyssmal performance must not produce correct builds!

Naturally, after spending nearly a decade developing a high-performance replacement for GNU make, I couldn’t let that statement stand. This post is probably going to be on the long side, so here’s the tl;dr summary:

  • You can guarantee correct builds with make, provided you follow best practices.
  • The worst-case runtime of any build tool if, of course, O(n2), but most, if not all, builds can be handled in O(n) time, without sacrificing correctness.
  • SCons’ performance problem is caused by design and implementation decisions in SCons, not some pathology of build structure.

What is required to ensure a correct build?

One of the fundamental tenents of the pro-SCons mythos is the idea that it is unique in its ability to guarantee correct builds. In reality, SCons is not doing anything particularly special in this regard. It’s true that by virtue of its design SCons makes it easier to get it right, but there’s nothing keeping you from enjoying the same assurances in make.

First: what is a correct build? Simply put, a correct build is one in which everything that ought to be built, is built. Note that by definition, a from-scratch build is correct, since everything is built in that case. So the question of “correct” or “incorrect” is really only relevant in regards to incremental builds.

So, what do we need in order to ensure a correct incremental build? Only three things, actually:

  1. A single, full-build dependency graph.
  2. Complete dependency information for every generated file.
  3. A reliable way to determine if a file is up-to-date relative to its inputs.

What SCons has done is made it more-or-less impossible, by design, to not have these three things. There is no concept like recursive make in the SCons world, so the only option is a single, full-build dependency graph. Likewise, SCons automatically scans input files in several programming languages to find dependency information. Finally, SCons uses MD5 checksums for the up-to-date check, which is a pretty darn reliable way to verify whether a given file needs to be rebuilt.

But the truth is, you can guarantee correct builds with make — you just have to adhere to long-standing best practices for make. First, you have to avoid using recursive make. Then, you need to add automatic dependency generation. The only thing that’s a little tricky is the up-to-date check: make is hardwired to use file timestamps, which can be spoofed, deliberately or accidentally — although to be fair, in most cases, timestamps are perfectly adequate. But even here, there’s a way out. You can use a smarter version of make that has a more sophisticated up-to-date mechanism, like ElectricMake or ClearMake. You can even shoehorn MD5 checksums into GNU make, if you like.

I can’t deny that SCons has made it easier to get correct builds. But the notion that it can’t be done with make is simply absurd.

What is the cost of a correct build?

Now we turn to the question of the cost of ensuring correctness. At its core, any build tool is just a collection of graph algorithms — first contructing the dependency graph, then traversing it to find and update out-of-date files. These algorithms have well-understood complexity, typically given as O(n + e), where n is the number of nodes in the graph, and e is the number of edges. It turns out that e is actually the dominant factor here, since it is at least equal to n, and at worst as much as n2. That means we can simplify the complexity to O(n + n2), or just O(n2).

Does this absolve SCons of its performance sins? Unfortunately it does not, because O(n2) is a worst-case bound — you should only expect O(n2) behavior if you’ve got a build that has dependencies between every pair of files. Think about that for a second. A dependency between every. pair. of. files. Here’s what that would look like in makefile syntax:

all: foo bar foo.c bar.c foo.h bar.h
foo:     bar foo.c bar.c foo.h bar.h
bar:         foo.c bar.c foo.h bar.h
foo.c:             bar.c foo.h bar.h
bar.c:                   foo.h bar.h
foo.h:                         bar.h

It’s ridiculous, right? I don’t know about you, but I’ve certainly never seen a build that does anything even remotely like that. In particular, the builds I used in my benchmarks don’t look like that. Fortunately, those builds are small and simple enough that we can directly count the number of edges in the dependency graph. For example, the smallest build in my tests consisted of:

2,000 C sources
+ 2,004 headers
+ 2,000 objects
+ 101 libraries
+ 100 executables

6,205 total files

So we have about 6,000 nodes in the graph, but how many edges does the graph contain? Lucky for us, SCons will print the complete dependency graph if we invoke it with scons –tree=all:

+-.
  +-SConstruct
  +-d1_0
  | +-d1_0/SConstruct
  | +-d1_0/f00000_sconsbld_d1_0
  | | +-d1_0/f00000_sconsbld_d1_0.o
  | | | +-d1_0/f00000_sconsbld_d1_0.c
  | | | +-d1_0/lup001_sconsbld_d1_0/f00000_sconsbld_d1_0.h
  ...

The raw listing contains about 35,000 lines of text, but that includes duplicates and non-dependency information like filesystem structure. Filter that stuff out and you can see the graph contains only about 12,000 dependencies. That’s a far cry from the 1,800,000 or so you would expect if this truly were a “worst-case” build. It’s clear, in fact, that the number of edges is best described as O(n).

Although I don’t know how (or even if it’s possible) to prove that this is the general case, it does make a certain intuitive sense: far from being strongly-connected, most of the nodes in a build dependency graph have just one or two edges. Each C source file, for example, has just one outgoing edge, to the object file generated from that source. Each object file has just one outgoing edge too, to the library or executable the object is part of. Sure, libraries and headers probably have more edges, since they are used by multiple executables or objects, but the majority of the stuff in the graph is going to fall into the “small handful of edges” category.

Now, here’s the $64,000 question: if the algorithms in a build tool scale in proportion to the number of edges in the dependency graph, and we’ve just shown that the dependency graph in question has O(n) edges, why does SCons use O(n2) time to execute the build?

Why is SCons so slow?

SCons’ O(n2) performance stems from its graph traversal implementation. Essentially, SCons scans the entire dependency graph each time it is looking for a file to update. n scans of a graph with O(n) nodes and edges equals an O(n2) graph traversal. There’s no mystery here. In fact, the SCons developers are clearly aware of this deficiency, as described on their wiki:

It’s worth noting that the Jobs module calls the Taskmaster once for each node to be processed (i.e., it’s O(n)) and the Taskmaster has an amortized performance of O(n) each time it’s called. Thus, the overall time is O(n^2).

But despite recognizing this flaw, they severely misjudged its impact, because they go on to state that it requires a “pathological” dependency graph in order to elicit this worst-case behavior from SCons. As we’ve shown here and in previous posts, even a terribly mundane dependency graph elicits O(n2) behavior from SCons. I shudder to think what SCons would do with a truly pathological dependency graph!

Obviously the next question is: why does SCons do this? That’s not quite as easy for me to explain, as an outside observer. To the best of my understanding, they rescan the graph just in case new dependencies are added to the dependency graph while evaluating a node in the graph — remember, in SCons the commands to update a file are expressed in Python, so they can easily manipulate the dependency graph even while the build is running.

Is it really necessary to rescan the dependency graph over and over? I don’t think so. In fact, make is proof that it is not necessary. I think there are two ways that SCons could address this problem: first, it could adopt GNU make’s convention of partitioning the build into distinct phases, one that updates dependency information, and a second that actually executes the build. In GNU make, that strategy allows for the introduction of new dependency information, while imposing only a one-time O(n) cost for restarting the make process if any new dependencies are found.

Alternatively, SCons could probably be made smarter about when a full rescan is required. Most of the time, even if new dependencies are added to the graph, they are added to the node being evaluated, not to nodes that were already visited. That is, when you scan a source file for implicit dependencies, you find the dependencies for that file not for other files in the build (duh). So most of the time, a full rescan is massive overkill.

The final word…?

Hopefully this is my last post on the subject of SCons performance. It is clear to me that SCons does not scale to large projects, and that the problem stems from design and implementation decisions in SCons, rather than some pathology in the build itself. You can get comparable guarantees of correctness from make, if you’re willing to invest the time to do things the right way. The payoff is a build system that is not only correct but has vastly better performance than SCons as your project grows. Why wouldn’t you want that?

%d bloggers like this: