Hi Chris,<div>what kind of application is that? At work I'm working on a ten years old production system with Oracle DBMS and it's 1 GB of size.<br><br></div><div>Thanks for share your analysis again,</div><div>Facu</div>
<div><br><div class="gmail_quote">On Wed, Nov 24, 2010 at 2:38 PM, Chris Muller <span dir="ltr"><<a href="mailto:ma.chris.m@gmail.com">ma.chris.m@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
There has been interest in Magma's read and query performance lately,<br>
so I thought I would share results of a recent benchmark test.<br>
<br>
It's an actual application which does very heavy reading and writing<br>
to a Magma repository. This test was 24GB of repository work, over<br>
two days. My goal was to determine, once the persistent model becomes<br>
many times the size of RAM and HD access appeared to become the<br>
limiting factor:<br>
<br>
- how fast is Magma, in terms of "objects per second" and in<br>
"kilobytes per second"?<br>
- how fast is this relative to the speed when the repository was empty?<br>
<br>
Conclusion:<br>
<br>
- Magma started at 4K objects per second (empty repository), 282K per second.<br>
- Finished with 2.3K objects per second (6GB repository), 750K per second.<br>
- Memory consumption by the image never exceeded 300MB.<br>
<br>
It is important to note, these times are from the client MagmaSession<br>
point-of-view, including the full server roundtrip plus<br>
materialization. Also, as can be seen from the attached data, there<br>
were many requests which only brought back one or two objects which,<br>
while dramatically lowering the overall reported throughput, is<br>
actually a real-world scenario for applications.<br>
<br>
Verbose description:<br>
<br>
When the test first started the repository was tiny, and there were<br>
only ~4K server requests during a 5-minute monitored period. The<br>
total number of objects read across all ~4K requests was ~17M for an<br>
average throughput of 4K objects per second (ops). As the model grew<br>
in size and complexity, more and more objects during the 5-minute<br>
monitored intervals were required to perform application-posting of<br>
the same number of input records; doubling to ~8K server requests<br>
during a 5-minute period, however only a few more objects brought back<br>
(total 21M), for an average of 2500 ops. 32 hours after that, the<br>
rate of reads dropped to about 2300 ops. This is due to two factors:<br>
<br>
- the objects were larger (e.g., more pointers to other objects)<br>
- they were less clustered (having been replaced with objects which<br>
could not be co-located with the original object)<br>
<br>
<br>
<br>
The fact that objects got bigger (e.g., more pointers to other<br>
objects) is corroborated by another stat, "the number of *bytes* per<br>
second" read off the HD in order to access those objects. Initially,<br>
when the test first started, there were about 4K *requests* for<br>
objects in one of the first five-minute periods, which were read at a<br>
rate of 282K per second. 24 hours later, there were only 2 times as<br>
many requests for objects, but were read-and-materialized at a rate of<br>
750K bytes / second.<br>
<font color="#888888"><br>
<br>
- Chris<br>
</font><br>_______________________________________________<br>
Magma mailing list<br>
<a href="mailto:Magma@lists.squeakfoundation.org">Magma@lists.squeakfoundation.org</a><br>
<a href="http://lists.squeakfoundation.org/mailman/listinfo/magma" target="_blank">http://lists.squeakfoundation.org/mailman/listinfo/magma</a><br>
<br></blockquote></div><br></div>