[
https://issues.apache.org/activemq/browse/AMQCPP-93?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_39480
]
Gerald Kaas commented on AMQCPP-93:
-----------------------------------
I've been spending the last day or so using a profiler against a test harness
we have which sends and receives within the same process 2000 BytesMessages of
approximately 500k in size with integers, doubles, strings, etc embedded within
the body of the message. Normally this task takes 5-7 seconds on a couple of
other middeware products we use but with the ActiveMQ C++ interface it was
taking 26 seconds or so. Essentially I made two rather small changes which cut
the time down to 15 seconds, which is approximately a 30% increase in
performance. Both of these changes were done within the DataInputStream.cpp
class. First, majority of the unmarshalling code was calling readFully for 1
byte. For the single byte extraction, I changed these readFully operations to
inputStream->read() which only return a single byte. There is an incorrectness
in the code that assumes that the inputStream->read(buffer, size) operation
will return -1 when inputStream is at its end. This is incorrect. As far as I
can tell, read(buffer, size) always throws an exception at the end rather than
returning -1. The second enhancement is I buffered up 256 characters before
appending the string in readString. Appending one character at a time is very
inefficient since it constantly needs to determine if the string object needs
to grow, realloc, copy the characters, etc.
The readString operation is still a major hitter in my profiling. It is calling
inputStream->read() millions of times which is a virtual function. Most strings
are tens, hundreds, or thousands of characters in size. Since it is virtual, it
cannot be inlined and they are CPU intensive when called repeatively. I'm sure
CPU usage could be dropped another 30-50% if we could move the readString
operation from the DataInputStream class to each inputStream class. That way
the virtual function is called only once for each string unmarshalling and more
inlining can happen. I'll leave it up to the experts in this group on how they
want to proceed.
Everything else seems fairly efficient, at least in the test case I was dealing
with. The only other item that showed up in the radar is the
ByteArrayOutputStream::write(unsigned char c) method. Again in this case, it is
constantly realloc, copying, etc when the vector needs to resize. There is a
setBuffer method in place but I need to evaluate it a little further to see if
we can preallocate our own buffer based on the size we already know what we
need up front.
Attached is my diffs. Let me know what you think.
> Performance analysis
> --------------------
>
> Key: AMQCPP-93
> URL: https://issues.apache.org/activemq/browse/AMQCPP-93
> Project: ActiveMQ C++ Client
> Issue Type: Task
> Affects Versions: 2.0
> Reporter: Nathan Mittler
> Assignee: Nathan Mittler
> Fix For: 2.2
>
> Attachments: amqcpp-perf1.patch, amqcpp-perf1v2.patch, bench1.cpp
>
>
> Do a performance analysis on openwire vs stomp.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
