Hi,
Before some major changes in our infrastructure, I was interested in the performance impact of moving indexes to our NAS (Emc Celerra v5.5). So I downloaded both dovecot-1.1.19 and dovecot-1.2.6, compiled them and wrote 2 confs for each of them, one with local indexes and another with indexes in the NAS.
Then I created 19 pam users, and tested system performance this way:
- uninstall current installed version
- install a new version
- start dovecot
- for each user run: imaptest user=user pass=pass host=localhost port=143 mbox=${MBOX} seed=123 secs=300 | tail -1 > file
- stop dovecot
I'm not sure if these tests are real, but our results are:
- In our conf, 1.1.19 and 1.2.6 performance are roughly the same
- Moving the indexes to our NAS makes the system 3 times slower
As we currently only have 1 system accessing mails or indexes, I tried later with nfs_* = no, and mmap_disable = no. This makes performance 10% better in any situation.
Our testing setup (very close to the real one)
- Dell PowerEdge 2850, 2 P4 Xeon 3GHz, 3GB RAM
- RHEL 4.7 (not sure about this, the last RHEL4) 32 bits
- mailboxes are always in a nfs filesystem
- for testing mbox file I've used dovecot-cvs.mbox
I've attached a couple of graphs to show this results (I've a lot of them, but result are pretty close from one to another). Also, dovecot -n output for 1.2.6 with indexes in the NAS. Only needed changes have been applied from one conf to another.
If you feel that the benchmark can be more "real" with some changes, just tell me. Also, if you are interested in the rest of graphs or date, tell me.
Next week I'll be quite busy, but as soon as possible I'll test with an iSCSI disk for indexes.
Aaaaaaaaaaagur.
Joseba Torre. Vicegerencia de TICs, área de Explotación
