> What are your thoughts on the right way to use SSDs in a RAID to
> enhance postgres I/O performance? In an earlier reply, you
> indicated one of a "RAID1+0 consisting of several spindles,
> NVRAM-based solution (SSD or PCIe card), or a SAN"
Well, it's a tiered approach. If you can identify your tables with the most critical OLTP needs, you can create a
separatetablespace specifically for SSD storage to give them the performance they need. After that, you might consider
partitioningeven those tables, as older data won't be accessed as often, so won't need those kind of IOPS long-term.
Olderpartitions could be slated toward the RAID.
Regarding what kind of SSD, just make sure the drives themselves are capacitor-backed. Current SSDs have only a few
microsecondsof write delay, but it's enough of a race condition to lead to corruption in power outages without some
assurancein-transit data is committed.
If you have the money, stuff like FusionIO PCIe cards are extremely fast, on the order of 10x faster than a standard
SSD.I'd personally reserve these for performance-critical things like online trading platforms, since they're so
costly.
Then of course, SANs can mix the world of RAID and SSD, because they often have internal mechanisms to deliver
requestedIOPS by spreading storage allocations along installed components necessary to match them. This is probably the
mostexpensive overall approach, but many larger companies either already have SANs, or will eventually need one anyway.
That's just a bird's eye view of everything. There's obviously more involved. :)
______________________________________________
See http://www.peak6.com/email_disclaimer/ for terms and conditions related to this email
