resizable_memory is slow to grow

[stkenned-havok]

Repro: create a large reserved region. Grow the committed region by a small amount each time.

Expected: time is proportional to the size delta.
Actual: time is proportional to the current total size

Seems to be related to passing the entire (0, newsize) to VirtualAlloc, giving quadratic resizing time.

[stkenned-havok]

Test I used
ConsoleApplication1.txt

[pknowles]

Hi @stephenatwork. Thanks very much for reporting this. And the neat example showing what should be happening. IIUC there's a few takeaways here...

1. I should be mapping only the addition pages needed to grow the current region (look like I'm already doing this in the linux implementation)
2. I can probably use VirtualAlloc rather than NtAllocateVirtualMemory for the windows resizable_memory implementation, which could skip loading the dll

Slightly related, I have some basic perf tests here but I should add some windows benchmarks too.

[pknowles]

Thanks again for reporting! Marking as closed with #11.

I created some new benchmarks for this in decodeless/benchmark#2...

Base

Easy to see with even 1MB when there's no exponential backoff 😳🙂.

This is up to 25MB, growing by 1009 bytes each iteration (e.g. to catch page boundary issues with the wrapper):

relative	ns/op	op/s	err%	total	benchmark
100.0%	5,304,335.79	188.53	0.1%	5.82	VirtualAlloc
0.6%	920,824,466.67	1.09	0.4%	31.36	resizable_memory (NtAllocateVirtualMemory)

Same but growing by the page size (4096):

relative	ns/op	op/s	err%	total	benchmark
100.0%	5,236,318.37	190.97	0.2%	6.05	VirtualAlloc
2.3%	225,346,000.00	4.44	0.1%	7.68	resizable_memory (NtAllocateVirtualMemory)

Mapping only the additional pages

• commit only delta pages when resizing mapped memory on windows #11

relative	ns/op	op/s	err%	total	benchmark
100.0%	5,023,913.59	199.05	0.2%	5.80	VirtualAlloc
98.7%	5,089,947.17	196.47	0.1%	6.06	resizable_memory (NtAllocateVirtualMemory)

relative	ns/op	op/s	err%	total	benchmark
100.0%	4,979,665.83	200.82	0.1%	6.05	VirtualAlloc
99.6%	5,000,425.23	199.98	0.1%	6.04	resizable_memory (NtAllocateVirtualMemory)

Using VirtualAlloc

Not much difference (even in 1MB cases). If any it'd be dynamically loading function pointers from ntdll.dll. I'll happily submit as it's much simpler :)

• VirtualAlloc instead of NtAllocateVirtualMemory #12

relative	ns/op	op/s	err%	total	benchmark
100.0%	5,057,159.46	197.74	0.5%	5.91	VirtualAlloc
99.8%	5,065,342.61	197.42	0.1%	6.03	resizable_memory (VirtualAlloc)

relative	ns/op	op/s	err%	total	benchmark
100.0%	4,986,085.09	200.56	0.1%	6.04	VirtualAlloc
99.8%	4,993,957.98	200.24	0.5%	6.08	resizable_memory (VirtualAlloc)

[stkenned-havok]

Great result, thanks!