GeneralPurposeAllocator: Considerably improve worst case performance

[squeek502]

Before this PR, GeneralPurposeAllocator could run into incredibly degraded performance in scenarios where the bucket count for a particular size class grew to be large. For example, if exactly slot_count allocations of a single size class were performed and then all of them were freed except one, then the bucket for those allocations would have to be kept around indefinitely. If that pattern of allocation were done over and over, then the bucket list for that size class could grow incredibly large, and to find a particular bucket, the entire (doubly linked) list would have to be scanned linearly.

This allocation pattern has been seen in the wild: Vexu/arocc#508 (comment)

In that case, the length of the bucket list for the 128 size class would grow to tens of thousands of buckets and cause Debug runtime to balloon to ~8 minutes whereas with the c_allocator the Debug runtime would be ~3 seconds.

To address this, there are three different changes happening here:

1. std.Treap is used instead of a doubly linked list for the lists of buckets. This takes the time complexity of searchBucket [used in resize and free] from O(n) to O(log n), but increases the time complexity of insert from O(1) to O(log n) [before, all new buckets would get added to the head of the list]. This is still a huge win because search happens way more often than insertion of new buckets. Note: Any data structure with O(log n) or better search/insert/delete would also work for this use-case.
2. If the 'current' bucket for a size class is full, the list of buckets is never traversed and instead a new bucket is allocated. Previously, traversing the bucket list could only find a non-full bucket in specific circumstances, and only because of a separate optimization that is no longer needed (before, after any resize/free, the affected bucket would be moved to the head of the bucket list to allow searchBucket to perform better on average). Now, the current_bucket for each size class only changes when either (1) the current bucket is emptied/freed, or (2) a new bucket is allocated (due to the current bucket being full or null). Because each bucket's alloc_cursor only moves forward (i.e. slots within a bucket are never re-used), we can therefore always know that any bucket besides the current_bucket will be full, so traversing the list in the hopes of finding an existing non-full bucket is entirely pointless.
3. Size + alignment information for small allocations has been moved into the Bucket data instead of keeping it in a separate HashMap. This offers an improvement over the HashMap since whenever we need to get/modify the length/alignment of an allocation it's extremely likely we will already have calculated any bucket-related information necessary to get the data.

The first change is the most relevant and accounts for most of the benefit here. Also note that the overall functionality of GeneralPurposeAllocator is unchanged.

In the degraded arocc case, these changes bring Debug performance from ~8 minutes to ~20 seconds.

Benchmark 1: test-master.bat
Time (mean ± σ): 481.263 s ± 5.440 s [User: 479.159 s, System: 1.937 s]
Range (min … max): 477.416 s … 485.109 s 2 runs
Benchmark 2: test-optim-treap.bat
Time (mean ± σ): 19.639 s ± 0.037 s [User: 18.183 s, System: 1.452 s]
Range (min … max): 19.613 s … 19.665 s 2 runs
Summary
'test-optim-treap.bat' ran
24.51 ± 0.28 times faster than 'test-master.bat'

Note: Much of the time taken on Windows in this particular case is related to gathering stack traces. With .stack_trace_frames = 0 the runtime goes down to 6.7 seconds, which is a little more than 2.5x slower compared to when the c_allocator is used.

These changes may or mat not introduce a slight performance regression in the average case:

Here's the standard library tests on Windows in Debug mode:

Benchmark 1 (10 runs): std-tests-master.exe
measurement mean ± σ min … max outliers delta
wall_time 16.0s ± 30.8ms 15.9s … 16.1s 1 (10%) 0%
peak_rss 42.8MB ± 8.24KB 42.8MB … 42.8MB 0 ( 0%) 0%
Benchmark 2 (10 runs): std-tests-optim-treap.exe
measurement mean ± σ min … max outliers delta
wall_time 16.2s ± 37.6ms 16.1s … 16.3s 0 ( 0%) 💩+ 1.3% ± 0.2%
peak_rss 42.8MB ± 5.18KB 42.8MB … 42.8MB 0 ( 0%) + 0.1% ± 0.0%

And on Linux:

Benchmark 1: ./test-master
Time (mean ± σ): 16.091 s ± 0.088 s [User: 15.856 s, System: 0.453 s]
Range (min … max): 15.870 s … 16.166 s 10 runs
Benchmark 2: ./test-optim-treap
Time (mean ± σ): 16.028 s ± 0.325 s [User: 15.755 s, System: 0.492 s]
Range (min … max): 15.735 s … 16.709 s 10 runs
Summary
'./test-optim-treap' ran
1.00 ± 0.02 times faster than './test-master'

---

Here are some more benchmark results using a very targeted benchmark that intentionally only does worst-case allocation patterns:

Benchmark code

conststd=@import("std");
pubfnmain() !void {
vargpa=std.heap.GeneralPurposeAllocator(.{}){};
deferstd.debug.assert(gpa.deinit() ==.ok);
constallocator=gpa.allocator();
constalloc_size=128;
constslot_count=@divExact(std.mem.page_size, std.math.ceilPowerOfTwoAssert(usize, alloc_size));
constrounds=5000;
varunfreed_slices: [rounds][]u8=undefined;
vari: usize=0;
while (i<rounds) : (i+=1) {
unfreed_slices[i] =tryallocator.alloc(u8, alloc_size);
for (0..(slot_count-1)) |_| {
constslice=tryallocator.alloc(u8, alloc_size);
allocator.free(slice);
}
}
for (&unfreed_slices) |slice| {
allocator.free(slice);
}
}

On Linux:

Debug:

Benchmark 1 (3 runs): ./gpa-degen-master
measurement mean ± σ min … max outliers delta
wall_time 3.66s ± 65.5ms 3.59s … 3.70s 0 ( 0%) 0%
peak_rss 41.9MB ± 2.36KB 41.9MB … 41.9MB 0 ( 0%) 0%
cpu_cycles 14.5G ± 280M 14.2G … 14.6G 0 ( 0%) 0%
instructions 25.1G ± 559M 24.5G … 25.5G 0 ( 0%) 0%
cache_references 98.7M ± 518K 98.1M … 99.1M 0 ( 0%) 0%
cache_misses 13.6M ± 156K 13.4M … 13.7M 0 ( 0%) 0%
branch_misses 56.8M ± 1.64M 55.0M … 58.2M 0 ( 0%) 0%
Benchmark 2 (9 runs): ./gpa-degen-optim-treap
measurement mean ± σ min … max outliers delta
wall_time 617ms ± 5.66ms 607ms … 624ms 0 ( 0%) ⚡- 83.2% ± 1.2%
peak_rss 41.9MB ± 1.81KB 41.9MB … 41.9MB 2 (22%) - 0.0% ± 0.0%
cpu_cycles 1.73G ± 12.9M 1.71G … 1.75G 0 ( 0%) ⚡- 88.0% ± 1.3%
instructions 2.79G ± 12.1M 2.78G … 2.82G 0 ( 0%) ⚡- 88.9% ± 1.5%
cache_references 38.7M ± 502K 37.9M … 39.3M 0 ( 0%) ⚡- 60.8% ± 0.8%
cache_misses 195K ± 10.1K 184K … 215K 0 ( 0%) ⚡- 98.6% ± 0.8%
branch_misses 4.39M ± 25.9K 4.36M … 4.43M 0 ( 0%) ⚡- 92.3% ± 1.9%

ReleaseFast:

Benchmark 1 (27 runs): ./gpa-degen-master-release
measurement mean ± σ min … max outliers delta
wall_time 187ms ± 4.41ms 178ms … 195ms 0 ( 0%) 0%
peak_rss 20.7MB ± 2.22KB 20.7MB … 20.7MB 0 ( 0%) 0%
cpu_cycles 705M ± 13.2M 680M … 728M 0 ( 0%) 0%
instructions 115M ± 17.8 115M … 115M 1 ( 4%) 0%
cache_references 29.5M ± 230K 29.1M … 30.0M 0 ( 0%) 0%
cache_misses 12.8M ± 15.0K 12.8M … 12.8M 0 ( 0%) 0%
branch_misses 35.4K ± 338 35.2K … 36.4K 1 ( 4%) 0%
Benchmark 2 (195 runs): ./gpa-degen-optim-treap-release
measurement mean ± σ min … max outliers delta
wall_time 25.6ms ± 2.80ms 20.3ms … 31.9ms 0 ( 0%) ⚡- 86.3% ± 0.7%
peak_rss 20.9MB ± 2.05KB 20.9MB … 20.9MB 0 ( 0%) + 1.0% ± 0.0%
cpu_cycles 35.2M ± 2.55M 30.8M … 46.5M 2 ( 1%) ⚡- 95.0% ± 0.3%
instructions 40.3M ± 912K 38.3M … 43.4M 1 ( 1%) ⚡- 65.0% ± 0.3%
cache_references 1.23M ± 186K 851K … 1.78M 3 ( 2%) ⚡- 95.8% ± 0.3%
cache_misses 12.3K ± 508 11.2K … 14.1K 6 ( 3%) ⚡- 99.9% ± 0.0%
branch_misses 52.5K ± 2.95K 48.9K … 57.8K 0 ( 0%) 💩+ 48.3% ± 3.2%

On Windows:

Debug:

Benchmark 1 (3 runs): gpa-degen-master.exe
measurement mean ± σ min … max outliers delta
wall_time 4.47s ± 165ms 4.29s … 4.62s 0 ( 0%) 0%
peak_rss 44.5MB ± 2.36KB 44.5MB … 44.5MB 0 ( 0%) 0%
Benchmark 2 (9 runs): gpa-degen-optim-treap.exe
measurement mean ± σ min … max outliers delta
wall_time 562ms ± 3.21ms 557ms … 567ms 0 ( 0%) ⚡- 87.4% ± 2.5%
peak_rss 44.5MB ± 2.05KB 44.5MB … 44.5MB 0 ( 0%) - 0.0% ± 0.0%

ReleaseFast:

Benchmark 1 (9 runs): gpa-degen-master-release.exe
measurement mean ± σ min … max outliers delta
wall_time 564ms ± 44.9ms 497ms … 603ms 0 ( 0%) 0%
peak_rss 23.5MB ± 2.05KB 23.5MB … 23.5MB 0 ( 0%) 0%
Benchmark 2 (120 runs): gpa-degen-optim-treap-release.exe
measurement mean ± σ min … max outliers delta
wall_time 41.8ms ± 1.55ms 38.5ms … 46.1ms 0 ( 0%) ⚡- 92.6% ± 1.4%
peak_rss 23.7MB ± 18.6KB 23.7MB … 23.9MB 1 ( 1%) + 0.9% ± 0.1%

---

Various notes:

• A memory pool is used for the Treap.Nodes. This has two slightly weird things:
  • Because the GPA doesn't have an init function and is directly instantiated instead, the memory pool can't use backing_allocator and instead always uses the page_allocator
  • The number of allocated Nodes will always stay at the peak number of Nodes necessary, meaning that e.g. if a program needs 5000 buckets at one point, then all 5000 of those nodes will live for the rest of the program even if all memory in the buckets is freed (but those 5000 nodes will also be re-used whenever a new node is needed).
• I initially used a skip list implementation that I wrote for this because I wasn't aware of std.Treap, but std.Treap slightly outperformed it in my benchmarks and provides all the same benefits.

[andrewrk]

Fantastic work!

• Because the GPA doesn't have an init function and is directly instantiated instead, the memory pool can't use backing_allocator and instead always uses the page_allocator

I was already thinking before this that GPA should lose the feature of using backing_allocator and become hard-coded to always use OS calls directly. An allocator that wraps an existing one and tries to provide some kind of features on top of it could be interesting, but it's a bit of a different concern than what GPA aims to provide. So, this is a step in the right direction IMO.