map: avoid allocations during PerCPUMap batch lookups#1823
Conversation
aibor
force-pushed
the
feat-per-cpu-batch-zero-alloc
branch
from
251da1f to
b794663
Compare
|
IIRC this will only work for types which do not have any implicit padding like uint64. will still go via the slow path because there is a 4 byte trailer. This is because we don't want to allow reading into memory that is not "visible" in the go type system. Users can work around this with explicit padding. Do you still want to add this in that case? If yes please update the commit message to explain this and probably drop a comment in the source code as well. |
Extend the optimization from b89a4cb ("map: avoid allocations during batch lookup of common types") to BatchLookup with PerCPUMaps This means that types without implicit padding will not incur allocations. See 4609dc7 ("map: zero-allocation operations for common types") for a detailed description of the behavior. ``` goos: linux goarch: amd64 pkg: github.com/cilium/ebpf cpu: AMD Ryzen 5 5600G with Radeon Graphics │ old.txt │ new.txt │ │ sec/op │ sec/op vs base │ Iterate/PerCPUHash/BatchLookup-8 551.3µ ± 1% 129.7µ ± 1% -76.47% (p=0.000 n=10) Iterate/PerCPUHash/BatchLookupAndDelete-8 568.2µ ± 1% 148.4µ ± 1% -73.88% (p=0.000 n=10) geomean 559.7µ 138.7µ -75.21% │ old.txt │ new.txt │ │ B/op │ B/op vs base │ Iterate/PerCPUHash/BatchLookup-8 89709.0 ± 0% 136.0 ± 0% -99.85% (p=0.000 n=10) Iterate/PerCPUHash/BatchLookupAndDelete-8 89799.0 ± 0% 224.0 ± 1% -99.75% (p=0.000 n=10) geomean 87.65Ki 174.5 -99.81% │ old.txt │ new.txt │ │ allocs/op │ allocs/op vs base │ Iterate/PerCPUHash/BatchLookup-8 1006.000 ± 0% 5.000 ± 0% -99.50% (p=0.000 n=10) Iterate/PerCPUHash/BatchLookupAndDelete-8 1006.000 ± 0% 5.000 ± 0% -99.50% (p=0.000 n=10) geomean 1.006k 5.000 -99.50% ``` Signed-off-by: Tobias Böhm <[email protected]>
aibor
force-pushed
the
feat-per-cpu-batch-zero-alloc
branch
from
b794663 to
70075be
Compare
|
Yes, I am aware of the behavior and would like to have the zero-allocations behavior for types satisfying the requirements as described by you. I have updated the commit message and referenced the original commit that has a very detailed description of the behavior. How would you like to have it documented in the code? As far as I can see the behavior is not really documented for the Map methods directly yet, or have I missed it? |
|
Maybe having it in the commit message is fine then! |
2 participants
Extend the optimization for batch lookups done in b89a4cb to PerCPUMaps as well.
It adapts the flow of the single value maps. For passed values objects that can not be used as backing memory, the behavior is the same as before.