[rabbitmq-discuss] Simple benchmark and results

[David Glaubman]

Matthias,

I did some profiling of a simple example (basically the SendString example, 
with Publish called in a loop). I understand this is not a realistic
scenario, 
but it does give some sense of where Publish takes time against .Net.

Anyway, almost all of the time was spent in Binarywriter.Write(byte) called
from Frame.WriteTo.  Each Write(byte) was taking about 3-5 mu-sec. 

Since this was so, I looked at NetworkBinarywriter, 
since it calls BinaryWriter.Write(byte) 4x for every uint32 or int32 and 2X
for every uint16.

Using BinaryWriter as a guide, I turned the overrides of Write(int32) etc
to write to a byte array in the proper network order 
and then write the array to the underlying Stream.

Running against the original RabbitMQ.Client dll, doing 10K Publishes of a
24-char string to a remote server,
I got about 1.5K messages per second:
   C:\> SendMultiString.exe ca1tesla1 directexchange direct key
123456789012345678901234 10000
   10000 messages sent in in 7248 mSec

Buffering the int/short Writes in a modified version of NetworkBinaryWriter, 
I got around 7.5K messages per sec.:  (5X speedup)
   C:\> SendMultiString.exe ca1tesla1 directexchange direct key
123456789012345678901234 10000
   10000 messages sent in in 1391 mSec

Also turning on Nagle (setting TCP_NODELAY to false), got it up to 15K per
sec: (total speed up > 10X)
  C:\> SendMultiString.exe ca1tesla1 directexchange direct key
123456789012345678901234 10000
  10000 messages sent in in 676 mSec

Nagle enabled with the original NetworkBinaryWriter was slightly better (in
this case) than the modified NetworkBinarywriter: (about 8.5K)
  C:\> SendMultiString.exe ca1tesla1 directexchange direct key
123456789012345678901234 10000
  10000 messages sent in in 1136 mSec


I'm not proposing that either the NBW should be rewritten along the lines
above or that Nagle's algorithm be enabled on Windows.
But I do think this shows there is something going on with Microsoft's
Socket/TCPClient implementations and there may be real gains available in
writing a whole Frame at once 
to the socket stream (at least when using TCPClient on Windows).

regards,
David

http://www.nabble.com/file/p26016750/NetworkBinaryReader.cs
NetworkBinaryReader.cs 
http://www.nabble.com/file/p26016750/SendMultiString.cs SendMultiString.cs 
-- 
View this message in context: http://www.nabble.com/Simple-benchmark-and-results-tp25968960p26016750.html
Sent from the RabbitMQ mailing list archive at Nabble.com.