Performance Question [message #10968] |
Sat, 10 March 2007 19:53 |
dietmar
Messages: 54 Registered: March 2007
|
Member |
|
|
Hi all,
we have done some performance tests running our mail-gateway software on OpenVZ. OpenVZ is extremely fast compared to other solutions, but we still get a 20% performance loss compared to running native. Our benchmark put the server under extremely high load (50 postfix smtpd, 40 spamassassin, 40 postgres database connections, ...). The Benchmark measures the maximum throughput of mails (Mails/second).
Is that 20% loss normal, or is it possible to improve that?
- Dietmar
|
|
|
Re: Performance Question [message #10972 is a reply to message #10968] |
Sun, 11 March 2007 01:18 |
rickb
Messages: 368 Registered: October 2006
|
Senior Member |
|
|
20% performance loss using Openvz isn't normal. Think of openvz as resource containers with tools to view container stats on the HN, and when a PID is in the context, it has many restrictions to "stay" in its context.
Other, true and heavier virtualization architectures such as vmware, xen, etc introduce an interface for the virtualized kernel to talk to the underlying HN. This is where a performance loss is induced.
20% is extremely high, and very unexpected.
Rick
-------------
Common Terms I post with: http://wiki.openvz.org/Category:Definitions
UBC. Learn it, love it, live it: http://wiki.openvz.org/Proc/user_beancounters
|
|
|
|
|
|
|
|
Re: Performance Question [message #11044 is a reply to message #11040] |
Mon, 12 March 2007 13:13 |
dev
Messages: 1693 Registered: September 2005 Location: Moscow
|
Senior Member |
|
|
It would be great, since then we could try to reproduce.
dietmar wrote on Mon, 12 March 2007 15:45 |
Quote: |
- what environment do you refer to when refering to native? host system? Which OS/kernel etc.?
|
debian/2.6.18-3
Quote: |
- during the test you used most likely different directories for storing emails etc. The fact is that HDD speed varies very much for disk beginning and end.
So to make results comparable you should do the following:
a) create a separate partition of size which will be used totally up (plus some small reserve)
b) mount this partition to /var/spool when testing host environment and /vz/root/VEID/var/spool when testing VE.
This will guarantee that the same physical sectors of HDD will be used in both tests.
|
Do you think that explains 15-20% Performance loss? Anyways, I cant change that for several reasons.
|
It can account easily for almost double performance variations in some cases. Though I don't claim it is 100% the issue in your test case, just trying to provide more likely reasons for performance differences and which we can check for.
dietmar wrote on Mon, 12 March 2007 15:45 |
Quote: |
- set VE UBC setting to unlimited. This will prevent resource shortages during the test.
There can be other tunings done and checked for (e.g. mounting /vz with noatime), but w/o test case and test procedure it is hard to advise.
|
I already set UBC settings to unlimited, but still get 15% less performance (which is quite good compared to other VM technologies)
dietmar wrote on Mon, 12 March 2007 15:45 |
I'm pretty sure it can be optimized further out. It's just not that easy to find out the bottleneck remotely maybe it's possible to get an access to your testbed?
Do you use 100Mb or 1Gb link?
I wonder if someone else did performance tests on comparable systems (mail server, mail gateways) running on OpenVZ?
|
|
[Updated on: Mon, 12 March 2007 13:14] Report message to a moderator
|
|
|
|
|
|