[
https://issues.apache.org/jira/browse/THRIFT-5948?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18074538#comment-18074538
]
Dmytro Shteflyuk commented on THRIFT-5948:
------------------------------------------
While working on serialization depth limits improvement, noticed a quick
performance win in a hot path when writing binary strings to memory buffer or
socket. For strings that are already have {{ASCII-8BIT}} encoding, C extension
will call into Ruby, which will return the original string. Even more, Ruby
code would always duplicate the string if it is frozen, even when the encoding
is already binary, which is unnecessary.
Instead, this PR re-implements Ruby land's
{{Thrift::Bytes.force_binary_encoding}} in C, and corrects Ruby logic to not
duplicate strings unneceserily.
h2. Benchmark
Benchmarked using
{code:bash}
test/rb/benchmarks/protocol_benchmark.rb --large-runs 5 --small-runs 50000
{code}
h3. Accelerated binary scenarios only
|| Scenario || Before || After || Delta ||
| c binary write large (1MB) structure 5 times | 0.758s | 0.473s | -37.6% |
| c binary read large (1MB) structure 5 times | 0.546s | 0.544s | -0.2% |
| c binary write 50000 small structures | 0.372s | 0.236s | -36.4% |
| c binary read 50000 small structures | 0.228s | 0.224s | -1.4% |
| *c-only suite total* | *1.901s* | *1.484s* | *-21.9%* |
| *c-only write total* | *1.130s* | *0.709s* | *-37.2%* |
| *c-only read total* | *0.773s* | *0.769s* | *-0.5%* |
h3. Full suite summary
|| Metric || Before || After || Delta ||
| Full suite total | 35.912s | 33.222s | -7.5% |
| Write scenarios total | 16.282s | 13.473s | -17.3% |
| Read scenarios total | 19.635s | 19.579s | -0.3% |
h3. Selected write-path highlights
|| Scenario || Before || After || Delta ||
| ruby binary write large (1MB) structure 5 times | 1.286s | 0.999s | -22.3% |
| c binary write large (1MB) structure 5 times | 0.741s | 0.473s | -36.2% |
| ruby compact write large (1MB) structure 5 times | 0.760s | 0.505s | -33.7% |
| ruby json write large (1MB) structure 5 times | 6.387s | 5.366s | -16.0% |
| ruby binary write 50000 small structures | 0.611s | 0.486s | -20.5% |
| c binary write 50000 small structures | 0.365s | 0.237s | -35.0% |
| ruby compact write 50000 small structures | 0.377s | 0.258s | -31.5% |
| ruby json write 50000 small structures | 2.973s | 2.439s | -18.0% |
Read-side scenarios stayed essentially flat overall, while write-side
allocation counts remained unchanged, which is consistent with reducing
write-path overhead rather than reducing allocation volume.
h3. Serializer / Deserializer benchmark
This also brings a huge win for {{Thrift::Serializer}} on accelerated protocols
(compact and binary accelerated):
|| Operation || Before || After || Delta || Allocations Before || Allocations
After ||
| {{Thrift::Serializer}} with {{CompactProtocolFactory}} ({{50000}}
serializations) | 0.407s | 0.259s | -36.3% | 1,800,031 | 1,800,031 |
| {{Thrift::Deserializer}} with {{CompactProtocolFactory}} ({{50000}}
deserializations) | 0.251s | 0.252s | +0.8% | 850,007 | 850,007 |
> Reduce Ruby binary extension write-path overhead in native
> force_binary_encoding helper
> ---------------------------------------------------------------------------------------
>
> Key: THRIFT-5948
> URL: https://issues.apache.org/jira/browse/THRIFT-5948
> Project: Thrift
> Issue Type: Improvement
> Components: Ruby - Library
> Reporter: Dmytro Shteflyuk
> Assignee: Dmytro Shteflyuk
> Priority: Major
> Time Spent: 20m
> Remaining Estimate: 0h
>
> The Ruby native extension still routes {{force_binary_encoding}} through
> {{Thrift::Bytes.force_binary_encoding}} via {{rb_funcall}}, even when the
> caller is already in C. This adds Ruby dispatch overhead on hot write paths
> in the native transport/protocol code, especially
> {{MemoryBufferTransport#write}}, {{BinaryProtocolAccelerated#write_binary}},
> and {{CompactProtocol#write_binary}}.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
