-
Eric Myhre authored
There's actually a doozy of performance considerations in this commit. Verifying (and then correcting) that comment kicked off a surprisingly deep and demanding binge of research. (Some of the considerations are still only "considerations", unfortunately -- one key discovery is that (surprise!) conclusive choices require more info than microbenchmarks alone can yield.) First, the big picture: One of the things we need to be really careful about throughout a system like go-ipld-prime (where we're dealing with large amounts of serialization) is the cost of garbage collection. Since we're often inside of an application's "tight loop" or "hot loop" or whatever you prefer to call it, if we lean on the garbage collector too heavily... it's very, very likely to show up a system-wide impact. So, in essence, we want to call "malloc" less. This isn't always easy. Sometimes it's downright impossible: we're building large tree structures; it's flatly impossible to do this without allocating some memory. In other cases, there are avoidable things: and in particular, one common undesirable source of allocations comes from "autoboxing" around interfaces. (More specifically, the name of the enemy here will often show up on profiling reports as "runtime.convT2I".) Sometimes this one can be avoided; other times, not. Now, a more detailed picture: There are actually several functions in the runtime that relate to memory allocation and garbage collection, and the less we use any of them, the better; but also, they are not all created equal. These are the functions that are of interest: - runtime.convT2I / runtime.convT2E - runtime.newObject - runtime.writeBarrier / runtime.gcWriteBarrier - runtime.convTstring / etc Most of these functions call `runtime.mallocgc` internally, which is why they're worth of note. (writeBarrier/gcWriteBarrier are also noteworthy, but are different beasts.) Different kinds of implementations of something like `justString` will cause the generated assembly to contain calls to various combinations of these runtime functions when they're converted into a `Node`. These are the variations considered: - Variation 1: `type justString string`: results in `runtime.convTstring`. - Variation 2: `type justString struct { x string }`: results in `runtime.convT2I`. - Variation 3: as above, but returning `&justString{val}`: results in `runtime.newobject` *and* its friends `runtime.writeBarrier` and `runtime.gcWriteBarrier`. The actual performance of these... it depends. In microbenchmarks, I've found the above examples are roughly: - Variation 1: 23.9 ns/op 16 B/op 1 allocs/op - Variation 2: 31.1 ns/op 16 B/op 1 allocs/op - Variation 3: 23.0 ns/op 16 B/op 1 allocs/op So, a couple things about this surprised me; and a couple things I'm still pretty sure are just microbenchmarks being misleading. First of all: *all* of these call out to `mallocgc` internally. And so we see an alloc-per-op reported in all three of them. (This actually kinda bugs me, because I feel like we should be able to fit all the requisite information in the first case directly into the interface, which is, if I understand correctly, already always two words; and arguably, the compiler might be smart enough to do this in the second case as well. But I'm wrong about this, for whatever reason, so let's just accept this one and move along.) But they vary in time. Why? Variation 2 seems to stand out as slower. Interestingly, it turns out `convT2E` and `convT2I` are extra problematic because they involve a call of `typedmemmove` internally -- as a comment in the source says, there's both an allocation, a zeroing, and then a copy here (at least, as of go1.12); this is a big bummer. In addition, even before getting that deep, if you check out the disassembly of just our functions: for our second variation, as inlined into our microbenchmark, there are 9 instructions, plus 1 'CALL'; vs only 3+1 for the first variation. This memmove and extra instructions seems to be the explainer for why our second variation (`struct{string}`) is significantly (~8ns) slower. (And here I thought variation two would do well! A struct with one field is the same size as the field itself; a string is one word of pointer; and an interface has another word for type; and that's our two words, so it should all pack, and on the stack! Alas: no.) Now back up to Variation 1 (just a typedef of a string): this one invokes `runtime.convTstring`, and while that does invoke `mallocgc`, there's a detail about how that's interesting: it does it with an ask for a small number of bytes. Specifically, it asks for... well, `unsafe.Sizeof(string)`, so that varies by platform, but it's "tiny". What's "tiny" mean? `mallocgc` has a specific definition of this, and you can see it by grepping the runtime package source for "maxTinySize": it's 16 bytes. Things under this size get special treatment from a "tiny allocator"; this seems to be why `runtime.convTstring` is relatively advantaged. (You can see benchmarks relating to this in the runtime package itself: try `go test -run=x -bench=Malloc runtime`. There's a *huge* cliff between MallocLargeStruct versus the rest of its fellows.) Variation 3 also appears competitive. This one surprises me, and this is where I still feel like microbenchmarks must be hoodwinking. The use of `runtime.newobject` seems to hit the same corners as `runtime.convTstring` at runtime in our situation here: it's "tiny"; that's neat. More confusingly, though, `runtime.writeBarrier` and `runtime.gcWriteBarrier` *should* be (potentially) very high cost calls. And for some reason, they're not. This particular workload in the microbenchmark must just-so-happen to tickle things in such a way that these calls are free (literally; within noise levels), and I suspect that's a happy coincidence in the benchmark that won't at all hold in real usage -- as any amount of real memory contention appears, the costs of these gc-related calls can be expected to rise. I did a few more permutations upon Variations 2 and 3, just out of curiosity and for future reference, adding extra fields to see if any interesting step functions revel themselves. Here's what I found: - {str,int,int,int,int} is 48 bytes; &that allocs the same amount; in speed, & is faster; 33ns vs 42ns. - {str,int,int,int} is 48 bytes; &that allocs the same amount; in speed, & is faster; 32ns vs 42ns. - {str,int,int} is 32 bytes; &that allocs the same amount; in speed, & is faster; 32ns vs 39ns. - {str,int} is 32 bytes; &that allocs the same amount; in speed, & is faster; 31ns vs 38ns. - {str} is 16 bytes; &that allocs the same amount; in speed, & is faster; 24ns vs vs 32ns. Both rise in time cost as the struct grows, but the non-pointer variant grows faster, and it experiences a larger step of increase each time the size changes (which in turn steps because of alignment). The &{str} case is noticeably faster than the apparently linear progression that starts upon adding a second field; since we see the number 16 involved, it seems likely that this is the influence of the "tiny allocator" in action, and the rest of the values are linear relative to each other because they're all over the hump where the tiny allocator special path disengages. (One last note: there's also a condition about "noscan" which toggles the "tiny allocator", and I don't fully understand this detail. I'd have thought strings might count as a pointer, which would cause our Variation 3 to not pass the `t.kind&kindNoPointers` check; but the performance cliff observation described in the previous paragraph seems to empirically say we're not getting kicked out by "noscan". (Either that or there's some yet-other phenomenon I haven't sussed.)) (Okay, one even-laster note: in the course of diving around in the runtime malloc code, I found an interesting comment about using memory "from the P's arena" -- "P" being one of the letters used when talking about goroutine internals -- and I wonder if that contributes to our little mystery about how the `gcWriteBarrier` method seems so oddly low-cost in these microbenchmarks: perhaps per-thread arenas combined with lack of concurrency in the benchmark combined with quickly- and sequentially-freed allocations means any gcWriteBarrier is essentially reduced to nil. However, this is just a theory, and I won't claim to totally understand the implications of this; commenting on it here mostly to serve as a pointer to future reading.) --- Okay. So what comes of all this? - I have two choices: attempt to proceed further down a rabbithole of microbenchmarking and assembly-splunking (and next, I think, patching debug printfs into the compiler and runtime)... or, I can see that last one as a step too far for today, pull up, commit this, and return to this subject when there's better, less-toy usecases to test with. I think the latter is going to be more productive. - I'm going to use the castable variation here (Variation 1). This won't always be the correct choice: it only flies here because strings are immutable anyway, and because it's a generic storage implementation rather than having any possibility of additional constraints, adjuncts from the schema system, validators, etc; and therefore, I don't actually care if it's possible to cast things directly in and out of this type (since doing so can't break tree immutability, and it can't break any of those other contracts because there aren't any). - A dev readme file appears. It discusses what choices we might make for other cases in the future. It varies by go native kind; and may be different for codegen'd types vs general storage implementations. - Someday, I'd like to look at this even further. I have a persistent, nagging suspicion that it should be possible to make more steps in the direction of "zero cost abstractions" in this vicinity. However, such improvements would seem to be pretty deep in the compiler and runtime. Someday, perhaps; but today... I started this commit in search of a simple diff to a comment! Time to reel it in. Whew. Signed-off-by: Eric Myhre <hash@exultant.us>
d0ce3ded