Remove DIRTY NIF flag #44
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
https://www.erlang.org/doc/man/erl_nif.html documentation indicates the threshold for being a dirty NIF should be about 1ms. Generally compress/decompress is much quicker than this. Testing indicates a significant overhead from using this flag when compress/decompress is in fact very fast.
|
|
|
|
Using a dirty_nif for smaller objects has a performance penalty so: - by default don't use the dirty nif if compressing < 250KB or uncompressing < 50KB - allow application to override and specifically request either dirty or quick method.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Thank-you for providing this repo, I've been able to use it to test adding zstd compression to the leveled database (which is used as a storage backend for the Riak database). After running some experiments adding this zstd compression library, it has been observed that there appears to be a significant cost from the recent PR which set the
ERL_DIRTY_JOB_CPU_BOUNDon the compress and decompress NIF.See OpenRiak#1. In the leveled perf_SUITE ct test (which is broad database load test), there is a 10%-30% cost from running ZSTD with the flag. Even with a simple unit test there is a clear difference.
With
ERL_DIRTY_JOB_CPU_BOUND:Compression time ZLIB 1739 LZ4 476 ZSTD 306 Decompression time ZLIB 272 LZ4 88 ZSTD 140Without
ERL_DIRTY_JOB_CPU_BOUND:Compression time ZLIB 1877 LZ4 456 ZSTD 251 Decompression time ZLIB 284 LZ4 86 ZSTD 110Reading the nif guidelines, it talks of the threshold being 1ms before a dirty CPU flag is required. Given the overhead is the flag appropriate for standard compress/decompress function? Perhaps there should be a separate entry point if one is sure the compress/decompress time is within those bounds?
I have limited experience with NIFs, so this PR is provided as a suggestion, but I'm happy to accept that this might not be the right thing to do.