Skip to content

PerformanceUpdate.md

Latest commit

 

History

History
56 lines (35 loc) · 3.69 KB

PerformanceUpdate.md

File metadata and controls

56 lines (35 loc) · 3.69 KB

FsFIX Performance Compared to QuickFixN and QuickFixJ

FsFIX performance is compared with QuickFixN (C#) and QuickFixJ (Java). Performance is in terms of the time taken to write and read FIX messages to and from byte arrays. BenchmarkDotNet was used to run the FsFIX and QuickFixN benchmarks. JMH was used for QuickFixJ. The GitHub repo for QuickFixN and QuickFixN benchmark projects can be found here. The times are the average (mean) times in microseconds.

(the main FsFIX page is here)

Results

benchmark FsFIX (original) FsFIX (optimized) QuickFixN (C#) QuickFixJ (Java)
write MarketDataRequest 3.1550 us 2.5762 us 8.4914 us 1.459 us
write QuoteStatusRequest with group 3.7849 us 3.3569 us 10.6389 us 1.216 us
write QuoteStatusRequest with nested group 5.3959 us 4.5921 us 19.7647 us 3.094 us
write NewOrderMultiLeg 30.2887 us 20.0561 us 43.2407 us 5.880 us
read MarketDataRequest 4.4538 us 2.7945 us 12.2478 us 1.956 us
read QuoteStatusRequest with group 6.3069 us 4.9191 us 11.9883 us 1.910 us
read QuoteStatusRequest with nested groups 8.3278 us 6.3515 us 19.5073 us 3.596 us
read NewOrderMultileg 44.1178 us 31.0434 us 107.2047 us 14.456 us

FsFIX is faster than QuickFIXN and slower than QuickFixJ. It is surprising that QuickFIXN is some much slower than QuickFixJ, they seem to have a similar implementation. QuickFixNs performance is probably not an issue in the context in which it is used.

These figures can be compared to CoralFIX, a commercial high performance FIX engine. CoralFIX benchmarks ran on a much higher performance PC than my laptop, but it is still interesting to compare.

(re-ran the benchmark for the original unoptimized version of FsFIX, as the original "write QuoteStatusRequest with nested group" benchmark had used a non nested-group FIX message)

FsFIX optimizations

  • FsFIX indexes byte buffers containing a FIX message. This index now optimistically assumes fields are in the order specified in the FIX specification, so the next field to read is one after the previous.
  • Improved conversion of primitive types when writing to byte arrays, ToString() is faster than sprintf
  • non-allocating BytesToInt32 and BytesToUint32 conversions

There is still scope for making FsFIX faster, possibly as fast as QuickFixJ, but it is hard to be sure.

How to run the benchmarks

Build the projects, in release mode for FsFIX and QuickFixN. Shutdown all other running programs.

FsFIX

  1. open console at: \benchmark\bin\Release
  2. run: benchmark.exe

QuickFixN

  1. open console at: \benchmarkQFDotNet\benchmarkQFDotNet\bin\Release
  2. run: benchmarkQFDotNet.exe

QuickFixJ

  1. open console at: \benchmarkQFJava\target
  2. run: java -jar target\benchmark.jar

Summary

FsFIX is already faster than QuickFIXN before optimization, while having the benefit of better type checking. It may be possible to close the gap with QuickFixJ.