Thread: Hardware purchase question
We currently are utilizing postgresql on 2 servers with the following configuration: 2 - 2.4 Ghz Xeon processors 4GB ram 4 36gb 10000rpm scsi drives configured for raid 10 We started out with one server and as we became IO bound we added the second. We are currently considering purchasing another identical server to go along with these. In addition to this we are also considering a scsi attached storage device in the 10 - 14 drive range configured for raid 10 in place of the onboard 4 drives we currently have. Daily about 30% of our data gets updated with about 2% new data. Our query load is about 60% reads and 40% writes currently. My question is what type of performance gains can I expect on average from swapping from 4 disk raid 10 to 14 disk raid 10? Could I expect to see 40 - 50% better throughput. The servers listed above are the dell 2650's which have perc 3 controllers. I have seen on this list where they are know for not performing well. So any suggestions for an attached scsi device would be greatly appreciated. Also, any thoughts on fibre channel storage devices? Thank You, Bo Stewart
Bo, > 2 - 2.4 Ghz Xeon processors > 4GB ram > 4 36gb 10000rpm scsi drives configured for raid 10 Hopefully you've turned OFF hyperthreading? > gains can I expect on average from swapping from 4 disk raid 10 to 14 disk > raid 10? Could I expect to see 40 - 50% better throughput. This is so dependant on application design that I can't possibly estimate. One big gain area for you will be moving the database log (pg_xlog) to its own private disk resource (such as a raid-1 pair). In high-write enviroments, this can gain you 15% without changing anything else. > The servers listed above are the dell 2650's which have perc 3 > controllers. I have seen on this list where they are know for not > performing well. So any suggestions for an attached scsi device would be > greatly appreciated. Also, any thoughts on fibre channel storage devices? The 2650s don't perform well in a whole assortment of ways. This is why they are cheap. NetApps seem to be the current best in NAS/SAN storage, although many people like EMC. Stay away from Apple's XRaid, which is not designed for databases. -- --Josh Josh Berkus Aglio Database Solutions San Francisco
>>>>> "BS" == Bo Stewart <bstewart@marketingsolutionsinc.com> writes: BS> The servers listed above are the dell 2650's which have perc 3 BS> controllers. I have seen on this list where they are know for not BS> performing well. So any suggestions for an attached scsi device would BS> be greatly appreciated. Also, any thoughts on fibre channel storage BS> devices? I have a 2450 and a 2650 both of which are totally sucking IO wise. The 2650 has a PERC3 card (LSI based) and has one channel holding a mirrored pair for the pg_xlog and OS, and the other channel has 14 U320 disks in a RAID5. If I'm lucky, I'll get 30MB/s out of the disks. Normally it hovers at 5 or 6MB/s on the big RAID. I'm currently shopping for non-Dell hardware to replace it :-( However, I keep getting conflicting advice. My choices are along these lines: Dual Xeon 64bit with built-in 6-disk RAID10 or RAID5 (LSI RAID card) Dual Opteron 64bit with built-in 6-disk RAID10 or RAID5 (LSI RAID card) Dual Opteron 64bit with external RAID via fibre channel (eg, nstor) I'm sure any of these will whip the bottom off the Dell 2650, but which will be the fastest overall? No way to know without spending lots of money to test. :-( Dell claims their new 2750 will be faster, but they've lost the battle already, and won't commit to any performance numbers. Won't even give me a ballpark number. -- =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= Vivek Khera, Ph.D. Khera Communications, Inc. Internet: khera@kciLink.com Rockville, MD +1-301-869-4449 x806 AIM: vivekkhera Y!: vivek_khera http://www.khera.org/~vivek/
> However, I keep getting conflicting advice. My choices are along > these lines: > > Dual Xeon 64bit with built-in 6-disk RAID10 or RAID5 (LSI RAID card) > Dual Opteron 64bit with built-in 6-disk RAID10 or RAID5 (LSI RAID card) > Dual Opteron 64bit with external RAID via fibre channel (eg, nstor) An Opteron, properly tuned with PostgreSQL will always beat a Xeon in terms of raw cpu. RAID 10 will typically always outperform RAID 5 with the same HD config. Fibre channel in general will always beat a normal (especially an LSI) raid. Dell's suck for PostgreSQL. Sincerely, Joshua D. Drake > > I'm sure any of these will whip the bottom off the Dell 2650, but > which will be the fastest overall? No way to know without spending > lots of money to test. :-( > > Dell claims their new 2750 will be faster, but they've lost the battle > already, and won't commit to any performance numbers. Won't even give > me a ballpark number. > -- Command Prompt, Inc., home of PostgreSQL Replication, and plPHP. Postgresql support, programming shared hosting and dedicated hosting. +1-503-667-4564 - jd@commandprompt.com - http://www.commandprompt.com Mammoth PostgreSQL Replicator. Integrated Replication for PostgreSQL
Attachment
Vivek, > Dual Xeon 64bit with built-in 6-disk RAID10 or RAID5 (LSI RAID card) > Dual Opteron 64bit with built-in 6-disk RAID10 or RAID5 (LSI RAID card) > Dual Opteron 64bit with external RAID via fibre channel (eg, nstor) Opteron over Xeon, no question. Not only are the Opterons real-world-faster, they are less severely affected by the CS bug. > I'm sure any of these will whip the bottom off the Dell 2650, but > which will be the fastest overall? No way to know without spending > lots of money to test. :-( The SAN is going to be faster with a good SAN. That being said, I understand that "a good SAN" is something like a $30,000 NetApp; the less expensive SANs/NASes don't seem to be more than an external drive enclosure with a raid chip (e.g. Apple XRaid). But we saw even a less expensive/slower EMC machine improve performance just moving the pg_xlog off of the local PERC RAID 5 and onto the SAN. So this is probably a good way to go if you can afford it. -- Josh Berkus Aglio Database Solutions San Francisco
Joshua D. Drake wrote: > > An Opteron, properly tuned with PostgreSQL will always beat a Xeon > in terms of raw cpu. > > RAID 10 will typically always outperform RAID 5 with the same HD config. > > Fibre channel in general will always beat a normal (especially an LSI) > raid. > > Dell's suck for PostgreSQL. Does anyone have any OS recommendations/experiences for PostgreSQL on Opteron? Thanks, Andrew
On Mon, 13 Dec 2004 09:23:13 -0800, Joshua D. Drake <jd@commandprompt.com> wrote: > > RAID 10 will typically always outperform RAID 5 with the same HD config. Isn't RAID10 just RAID5 mirrored? How does that speed up performance? Or am I missing something? -- Mitch
Mitch Pirtle wrote: > On Mon, 13 Dec 2004 09:23:13 -0800, Joshua D. Drake > <jd@commandprompt.com> wrote: > >>RAID 10 will typically always outperform RAID 5 with the same HD config. > > > Isn't RAID10 just RAID5 mirrored? How does that speed up performance? > Or am I missing something? > > -- Mitch Hi Mitch, Nope, Raid 10 (one zero) is a mirror is stripes, no parity. with r10 you get the benefit of a full mirror which means your system does not need to calculate the XOR parity but you only get 50% disk usage. The mirror causes a slight write hit as the data needs to be split between two disk (or in this case, to striped pairs) but reads can be up to twice as fast (theoretically). By adding the stripe you negate the write hit and actually gain write performance because half the data goes to mirror A, half to mirror B (same with reads, roughly). Raid 10 is a popular choice for software raid because of the reduced overhead. Raid 5 on the otherhand does require that a parity bit is calculated for every N-1 disks. With r5 you get N-1 disk usage (you get the combined capacity of 3 disks in a 4 disk r5 array) and still get the benefit of striping across the disks so long as you have a dedicated raid asic that can do the XOR calculations. Without it, specially in a failure state, the performance can collapse as the CPU performs all that extra math. hth Madison
Madison Kelly wrote: > Nope, Raid 10 (one zero) is a mirror is stripes, no parity. with r10 Woops, that should be "mirror of stripes". By the way, what you are thinking of is possible, it would be 51 (five one; a raid 5 built on mirrors) or 15 (a mirror of raid 5 arrays). Always be careful, 10 and 01 are also not the same. You want to think carefully about what you want out of your array before building it. Madison
Madison Kelly <linux@alteeve.com> writes: > Without it, specially in a failure state, the performance can collapse as > the CPU performs all that extra math. It's really not the math that makes raid 5 hurt. It's that in order to calculate the checksum block the raid controller needs to read in the existing checksum block and write out the new version. So every write causes not just one drive seeking and writing, but a second drive seeking and performing a read and a write. The usual strategy for dealing with that is stuffing a huge nonvolatile cache in the controller so those reads are mostly cached and the extra writes don't saturate the i/o throughput. But those kinds of controllers are expensive and not an option for software raid. -- greg
You are right, I now remember that setup was originally called "RAID 10 plus 1", and I believe is was an incorrect statement from an overzealous salesman ;-) Thanks for the clarification! - Mitch On Mon, 03 Jan 2005 15:19:04 -0500, Madison Kelly <linux@alteeve.com> wrote: > Madison Kelly wrote: > > Nope, Raid 10 (one zero) is a mirror is stripes, no parity. with r10 > > Woops, that should be "mirror of stripes". > > By the way, what you are thinking of is possible, it would be 51 (five > one; a raid 5 built on mirrors) or 15 (a mirror of raid 5 arrays). > Always be careful, 10 and 01 are also not the same. You want to think > carefully about what you want out of your array before building it.
...and on Mon, Jan 03, 2005 at 03:44:44PM -0500, Mitch Pirtle used the keyboard:
>
> You are right, I now remember that setup was originally called "RAID
> 10 plus 1", and I believe is was an incorrect statement from an
> overzealous salesman ;-)
>
Just an afterthought - that could well be the unfortunate consequence of
salesmen specializing in sales as an act rather than the goods they were
selling - it might be that he/she was referring to the specifics of the
concrete configuration they were selling you (or trying to sell you),
which should, in the case you were mentioning, probably be called "a
RAID10 array with a hotspare drive" - that is, it would be preconfigured
to, upon the failure of one of array members, detect the failed drive and
automatically replace it with one that has been sitting there all the time,
doing nothing but waiting for one of its active companions to fail.
But this already falls into the category that has, so far, probably
caused the vast majority of misunderstandings, failed investments and
grey hair in RAID, namely data safety, and I don't feel particularly
qualified for getting into specifics of this at this moment, as it
happens to be 2AM, I had a couple of beers (my friend's birthday's due)
and I'm dying to get some sleep. :)
HTH, cheers,
--
Grega Bremec
gregab at p0f dot net