Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On 10/12/2011 11:17 AM, Kushnir, Michael (NIH/NLM/LHC) [C] wrote: > Good morning, > > I am running a two node DRBD 8.3.8 primary/primary cluster with GFS2 on RHEL 5.7. GFS2 is exported via GNBD. My underlying hardware is 2 x Dell C2100 server with LSI 9260-8i RAID controllers. RAID set is RAID10 with 10 x 1TB SATA 7.2k disks. I use 1MB stripe size as well as read and write caching. NIC is dual-port Myri 10GbE (10G-PCIE2-8B2-S2) connected to Cisco M4900. Ports are in LACP group. Red Hat cluster suite? If so, LACP isn't supported (only Active/Passive is for redundancy). This is aside your question though, of course. > My questions: > > 1. In the case of a 2-primary split brain (switch hiccup, etc), I would like server #1 to always remain primary and server #2 to always shut down. I would like this behavior because server #2 can't become secondary because GNBD is not going to release it. What is the best way to accomplish this? I'd suggest putting a delay in the second node's fence call. That way, in a true split brain, the primary will have a good head start in calling the fence against the backup node. However, delay to recovery when the primary really does fail will grow by the delay amount. > 2. I've tried the deadline queue manager as well as CFQ. I've noticed no difference. Can you please elaborate on why deadline is better, and how can I measure any performance difference between the two? I've not used GNDB, so I dare not speculate. > 3. It seems that GNBD is the biggest source of latency in my system. It decreases IOPS by over ~50% (based on DD tests compared to the same DRBD based GFS2 mounted locally). I've also tried Enterprise iSCSI target as an alternative and the results were not much better. The latency on my LAN is ~0.22ms. Can you offer any tuning tips? > > Thanks, > Michael There is overhead because of the distributed nature of clustered storage. However, I can't say where/why your latency is coming from so I don't have much to recommend at this time. If you create a simple DRBD resource and test, what is the overhead relative to the bare drives underneath? How does that change when you add simple GFS2? How about if you used CLVMd as a (test) alternative? If the latency is fairly close between GFS2 and clvmd, it's possibly DLM overhead. -- Digimer E-Mail: digimer at alteeve.com Freenode handle: digimer Papers and Projects: http://alteeve.com Node Assassin: http://nodeassassin.org "At what point did we forget that the Space Shuttle was, essentially, a program that strapped human beings to an explosion and tried to stab through the sky with fire and math?"