Dec 9, 2019 8 min read

Benchmark Surprises

I love a good mystery. If you really want my undivided attention, present me with a program which does a thing, and also a proof that it can’t possibly do that thing. For instance, if you can make a change to a program that can’t affect its performance, and change its performance dramatically? I’m hooked.

So let’s look at a few real-world examples of performance mysteries which showed up in benchmarks, and explanations of them. This might, if you’re a boring person, also be used as advice on avoiding these kinds of weird and magical behaviors. I guess. If that’s what you’re into.

All of these examples showed up in #performance in Gopher slack. I’m abbreviating and summarizing a bit for space, feel free to go skim the archives for more detail.

Also, for those who like spoilers: The answer is nearly always “the microbenchmark is not an accurate reflection of real workloads”, the interesting part is in trying to figure out how it’s not like real workloads.

Buffering for Performance

Someone came in with a fascincatingly weird behavior. They had some code which had a data structure which preallocated some fairly large buffers for performance reasons, but they didn’t think the buffers were really helping. So, they altered the code to not use the buffers, and that was fine. Then they removed the buffers, and suddenly got a 25% performance penalty.

Now, if ceasing to use the buffers cost you 25% of your performance, that could make sense; obviously in that case the buffers are useful. But in this case, the buffers weren’t being used at all. You could rename the fields, not change any code, and everything still compiles and runs.

There’s one option here for bisecting: Leave the buffers in, but don’t allocate them. Which is to say, separate “the slice field exists in the struct” from “the slice has been created with make(...)”. And that reveals that it’s making the slices that helps.

So, what’s happening? Go’s garbage collector tries pretty aggressively to keep the total allocated memory within about a factor of two of the size of the active heap. What this means is that, on a microbenchmark where there’s very little going on except the specific data structures being tested, the allocation of those large buffers meant the garbage collector was much less aggressive, and ran significantly less often.

In other words: The microbenchmark is not an accurate reflection of real workloads. In a real workload, there’s going to be a lot more other things on the heap, and thus the creation of a couple of buffers isn’t going to materially affect the garbage collector’s behavior.

Manual Loop Faster Than range Operator

I used to do a lot of C with primitive compilers, and I still tend to think “oh hey I bet I can do this faster” and be completely wrong, but one day, I thought “I wonder whether manually looping through this slice would be faster or slower than using the range operator”, so I made a test case and discovered that it was faster. Then I adjusted the test case a bit and discovered that it was slower. This went back and forth a few times.

The breakthrough happened when, being in a hurry, I made a stupid mistake, and ended up comparing the wrong two implementations: Specifically, instead of comparing two different implementations, I compared two copies of the same implementation. And discovered that one of them was well over 30% faster than the other.

Here’s a link to a version of this program on the Go Playground:

The relevant code:

func BenchmarkRange(b *testing.B) {
	for i := 0; i < b.N; i++ {
		RangeOp(a1, a2)
		RangeOp(a2, a1)

func BenchmarkHand(b *testing.B) {
	for i := 0; i < b.N; i++ {
		HandRange(a1, a2)
		HandRange(a2, a1)

The RangeOp and HandRange functions are different, but those differences turn out not to matter. Instead, the order of these functions matters; if I move the second one above the first in the source file, the performance changes. Sometimes. It depends on the exact day, the exact version of the compiler, and so on.

What’s actually happening? Well, honestly, I don’t know precisely. Running these under perf shows that they get different behavior for cache misses and loads, and sometimes slightly different behavior for branch prediction.

What appears to be at issue is, in effect, the addresses of these functions. A friend who has done CPU microarchitecture work at some length suggests that it’s possible this is hitting CPU cache associativity limits – if the code for one of these functions ends up in the same cache bucket as the code for the function it’s calling, they can evict each other from the cache sometimes, which could result in significantly lower performance. It could also be alignment of some type, or the exact ranges of jump instructions.

What’s important is to understand that you can get a 30% variance in performance for a hot path due to essentially unforseeable code layout changes. Oops.

Worse yet, in this case, that could affect a real workload, but there’s nothing you could really do about it. Any change elsewhere in the source file could affect this – adding a fmt.Printf call in a function that isn’t even called could, after all, change the generated code for the whole file.

But it’s still not a reflection of real workloads, in that the observed performance difference isn’t actually even related to the code being executed.

The Computer Can Press Its Own Turbo Button

Long ago, some software was dependent on CPU speed enough that early machines which could run at speeds other than the standard 4.77MHz often had a button to toggle whether the computer ran at some reasonably native clock speed, or at the much slower rate that the software had been written for.

Modern software is much less likely to fail completely when clockspeed changes, and we have clocks other than cycle counters, but when you’re benchmarking, it does matter what speed the CPU runs at.

Unfortunately, what speed the CPU runs at varies. A lot. On my laptop, for instance, the nominal clock speed is about 3GHz. If not much is happening, it is probably running around 1100MHz. But even talking about the speed of “the” CPU is misleading; despite this being a single physical chip, it has four cores, and 8 virtual CPUs thanks to hyperthreading. So let’s see what I’m running at under light load:

$ grep MHz /proc/cpuinfo
cpu MHz         : 2328.913
cpu MHz         : 2083.069
cpu MHz         : 2042.439
cpu MHz         : 2951.134
cpu MHz         : 2948.966
cpu MHz         : 1883.699
cpu MHz         : 1998.988
cpu MHz         : 2378.981

Note that there’s exactly one CPU reporting <1900MHz, despite the CPU having two virtual CPUs per core. Why? Who knows. But under these circumstances, the same code getting switched from one core to another could move it from a 1.9GHz CPU to a 2.9GHz CPU. That kind of difference could be fairly significant.

One partial solution, for x86 systems, is perflock, which does two things. First, it is a sort of performance-testing mutex; it won’t let two things that are using it run at once, so they won’t interfere with each other. (This won’t help as much if you’re also running CPU-intensive things, such as a text editor, or a text chat client, that can consume gigabytes of memory and entire CPUs.) Secondly, it will interact with the system’s power management and CPU performance stuff to try to lock a CPU at a fixed rate. Note that the default value may be higher than a system can sustain – you may want to run with -governor 70% or something similar to pick something that won’t cause thermal throttling, or else you still get unexpected performance variance.

In this case, the highly erratic behavior of CPU performance does reflect real workloads; modern systems have highly erratic behavior sometimes! Unfortunately, that’s useless when comparing the performance of different versions of code to see which is faster. It’s useful to do that testing in an environment which intentionally doesn’t represent real-world workloads, when you want reproducible data.

Computer Too Bored, Not Paying Attention

Someone was doing some latency tests on a service, using vegeta, and was getting really strange results. Intuitively, my expectation would be that in general, higher loads lead to more latency; there’s more probability that you are waiting for something else to complete, or contending for a lock. What they actually observed was dramatically lower latency at higher request rates.

Here’s the Github issue where this showed up.

In this case, the issue appears to be that, on modern systems, CPUs tend to very quickly go into idle states and lower their clock speed when there’s not enough workload to saturate them. They don’t immediately speed up fully again, so if you’re on a processor that’s nominally 3GHz, and it’s currently running at 800MHz, it’s going to be running at about 25% of nominal speed, and that means things will take 4x longer.

In experimentation, I found that having even one low-priority task on a machine that woke up every nanosecond or so kept the machine from doing this. I would summarize the problem as “the computer is bored and is not really paying attention”.

The problem here is that usually when we’re talking about performance, we’re thinking about amount of work done over time, or how long it takes per work item, and for that purpose, it doesn’t matter how fast things are when there aren’t nearly enough to saturate the CPU. However, latency of individual requests can also matter, and having latency suddenly get noticably worse when things are quiet is going to be a problem for some workloads. It isn’t totally solved by running a small amount of busywork to keep the CPU up, or using perflock, but that seems to address most of the gap.

Special thanks to:

  • The entire #performance channel.