We're seeing a *serious *performance issue with mod_jk and large (e.g.
500MB+) file transfers. [This is with Apache 2.0.55, Tomcat 5.0.30, and
various recent mod_jk including 1.2.20.]
The performance of downloading the file via Apache is good, as is the
performance when downloading directly from Tomcat. The performance when
downloading from Tomcat through Apache via mod_jk is, however, quite
abysmal. I'd obviously expect *some* degradation due to the extra
interprocess hop, but given that this is a just a single-user,
single-request test, I'd expect that the network would still be the
limiting factor -- or at least that the degradation would be in the
order of 25% of less. What we're seeing, however, is far worse:
On Windows:
* Apache 2.0.55, Tomcat 5.0.30, and mod_jk 1.2.20 - Started at
10 MB/sec ended at 3 MB/sec with mod_deflate disabled (1.5
MB/sec with mod_deflate enabled)
* Apache 2.0.55, Tomcat 5.0.30, and mod_jk 1.2.19 - Disabling
JkFlushPackets only slightly improved performance.
* Apache 2.2.3 with Tomcat 5.5.20 w/ the native connector -
Didn't work period. I didn't have a chance to look into it,
but the download failed after getting serveral packets (!)
* Apache 2.2.3 with Tomcat 5.5.20 w/o the native connector - Was
only slightly slower than going straight through Apache
about 7-8 MB/sec
On Solaris:
* Apache 2.0.55, Tomcat 5.0.30, recent mod_jk - Fairly constant
4MB/s when going through mod_jk, 10MB/s when just downloading
via Apache
[This issue originally was thought to be Windows specific, which is
why we have many more results for Windows.]
Obviously if our end goal was simple static file transfers we'd just
share/mirror them to Apache to solve this (we need the load balancing
flexibility, etc, of mod_jk, so directly using Tomcat is not really an
option -- nor is doing non-AJP-proxying). The static file case is the
simplified reproduction of our real issue, however, which is large file
downloads from our (Java-based) content store.
We had much better results with Apache 2.2.3 and Tomcat 5.5.20 with
tcnative, but we really don't want to force a move to 2.2.x and Tomcat
5.5.x in this case and we've had issues with tcnative (which we *hope*
may be resolved with 1.1.8). Overall we'd much prefer to get mod_jk
working reasonably than to force a disruptive move to 2.2.x right now.
Is this a known issue? Any pointers as to where/how to look for the
performance bottleneck? Some VTune examination showed that almost all
of Apache's CPU time during this time was in libapr.dll, but that's
obviously not terribly specific.
--
Jess Holle
