Re: [PERFORM] Planning a new server - help needed
Greg Smith wrote: As for SCSI vs. SATA, I collected up the usual arguments on both sides at http://www.postgresqldocs.org/index.php/SCSI_vs._IDE/SATA_Disks Why do you claim that 'More platters also means slower seeks and generally slower performance.'? On the face of it, it should mean that the number of track step operations is reduced, even if the drive doesn't buffer a slice of tracks aross all platters (which would help if it did). I'm not entirely sure why the extra platters should really count as more moving parts since I think the platter assembly and head assembly are both single parts in effect, albeit they will be more massive with more platters. I'm not sure how much extra bearing friction that will mean, but it is reasonable that some extra energy is going to be needed. Recent figures I've seen suggest that the increased storage density per platter, plus the extra number of platters, means that the streaming speed of good 7200rpm SATA drives is very close to that of good 15000rpm SAS drives - and you can choose which bit of the disk to use to reduce seek time and maximise transfer rate with the oversize drives. You can get about 100MB/s out of both technologies, streaming. It may be worth considering an alternative approach. I suspect that a god RAID1 or RAID1+0 is worthwhile for WAL, but you might consider a RAID1 of a pair of SSDs for data. They will use a lot of your budget, but the seek time is negligible so the effective usable performance is higher than you get with spinning disks - so you might trade a fancy controller with battery-backed cache for straight SSD. I haven't done this, so YMMV. But the prices are getting interesting for OLTP where most disks are massively oversized. The latest Samsung and SanDisk are expensive in the UK but the Transcend 16GB TS16GSSD25S-S SATA is about $300 equiv - it can do 'only' 'up to' 28MB/s write and you wouldn't want to put WAL on one, but sustaining 15-20MB/s through random access to a real disk isn't trivial. If average access is 10ms, and you write 100MB/s streaming, then you have to ask yourself if you going to do 80 or more seeks a second. James James -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Planning a new server - help needed
Why do you claim that 'More platters also means slower seeks and generally slower performance.'? More platters - more heads - heavier head assembly - slower seek time But.. More platters - higher density - less seek distance (in mm of head movement) - faster seek time As usual, no clear-cut case, a real-life test would tell more interesting things. I'm not entirely sure why the extra platters should really count as more moving parts since I think the platter assembly and head assembly are both single parts in effect, albeit they will be more massive with more platters. I'm not sure how much extra bearing friction that will mean, but it is reasonable that some extra energy is going to be needed. Since the bearings are only on one side of the axle (not both), a heavier platter assembly would put more stress on the bearing if the disk is subject to vibrations (like, all those RAID disks seeking together) which would perhaps shorten its life. Everything with conditionals of course ;) I remember reading a paper on vibration from many RAID disks somewhere a year or so ago, vibration from other disks seeking at the exact same time and in the same direction would cause resonances in the housing chassis and disturb the heads of disks, slightly worsening seek times and reliability. But, on the other hand, the 7 disks raided in my home storage server never complained, even though the $30 computer case vibrates all over the place when they seek. Perhaps if they were subject to 24/7 heavy torture, a heavier/better damped chassis would be a good investment. It may be worth considering an alternative approach. I suspect that a god RAID1 or RAID1+0 is worthwhile for WAL, but Actually, now that 8.3 can sync to disk every second instead of at every commit, I wonder, did someone do some enlightening benchmarks ? I remember benchmarking 8.2 on a forum style load and using a separate disk for WAL (SATA, write cache off) made a huge difference (as expected) versus one disk for everything (SATA, and write cache off). Postgres beat the crap out of MyISAM, lol. Seems like Postgres is one of the rare apps which gets faster and meaner with every release, instead of getting slower and more bloated like everyone else. Also, there is a thing called write barriers, which supposedly could be used to implement fsync-like behaviour without the penalty, if the disk, the OS, the controller, and the filesystem support it (that's a lot of ifs)... I haven't done this, so YMMV. But the prices are getting interesting for OLTP where most disks are massively oversized. The latest Samsung and SanDisk are expensive in the UK but the Transcend 16GB TS16GSSD25S-S SATA is about $300 equiv - it can do 'only' 'up to' 28MB/s write and Gigabyte should revamp their i-RAM to use ECC RAM of a larger capacity... and longer lasting battery backup... I wonder, how many write cycles those Flash drives can take before reliability becomes a problem... -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Planning a new server - help needed
PFC wrote: Why do you claim that 'More platters also means slower seeks and generally slower performance.'? More platters - more heads - heavier head assembly - slower seek time Note sure I've sen a lot of evidence of that in drive specifications! Gigabyte should revamp their i-RAM to use ECC RAM of a larger capacity... and longer lasting battery backup... You would think a decent capacitor or rechargable button battery would be enough to dump it to a flash memory. No problem with flash wear then. I wonder, how many write cycles those Flash drives can take before reliability becomes a problem... Hard to get data isn't it? I believe its hundreds of thousands to millions now. Now each record in most OLTP tables is rewritten a few times unless its stuff that can go into temp tables etc, which should be elsewhere. Index pages clearly get rewritten often. I suspect a mix of storage technologies will be handy for some time yet - WAL on disk, and temp tables on disk with no synchronous fsync requirement. I think life is about to get interesting in DBMS storage. All good for us users. James -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Planning a new server - help needed
On Sat, 29 Mar 2008, PFC wrote: Why do you claim that 'More platters also means slower seeks and generally slower performance.'? More platters - more heads - heavier head assembly - slower seek time I recall seeing many designs with more platters that have slower seek times in benchmarks, and I always presumed this was the reason. That's the basis for that comment. I'll disclaim that section a bit. Actually, now that 8.3 can sync to disk every second instead of at every commit, I wonder, did someone do some enlightening benchmarks ? I've seen some really heavy workloads where using async commit helped group commits in a larger batches usefully, but I personally haven't found it to be all that useful if you're already got a caching controller to accelerate writes on the kinds of hardware most people have. It's a great solution for situations without a usable write cache though. Also, there is a thing called write barriers, which supposedly could be used to implement fsync-like behaviour without the penalty, if the disk, the OS, the controller, and the filesystem support it (that's a lot of ifs)... The database can't use fsync-like behavior for the things it calls fsync for; it needs the full semantics. You're either doing the full operation, or you're cheating and it doesn't do what it's supposed to. Write barriers aren't any improvement over a good direct I/O sync write setup for the WAL. There may be some limited value to that approach for the database writes at checkpoint time, but again there's a real fsync coming at the end of that and it's not satisfied until everything is on disk (or in a good disk controller cache). Gigabyte should revamp their i-RAM to use ECC RAM of a larger capacity... and longer lasting battery backup... I saw a rumor somewhere that they were close to having a new version of that using DDR2 ready, which would make it pretty easy to have 8GB on there. I wonder, how many write cycles those Flash drives can take before reliability becomes a problem... The newer SSD drives with good write leveling should last at least as long as you'd expect a mechanical drive to, even in a WAL application. Lesser grades of flash used as disk could be a problem though. -- * Greg Smith [EMAIL PROTECTED] http://www.gregsmith.com Baltimore, MD -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Planning a new server - help needed
Greg Smith wrote: Gigabyte should revamp their i-RAM to use ECC RAM of a larger capacity... and longer lasting battery backup... I saw a rumor somewhere that they were close to having a new version of that using DDR2 ready, which would make it pretty easy to have 8GB on there. I'm hoping - probably in vain - that they'll also include a CF/SD slot or some onboard flash so it can dump its contents to flash using the battery backup. For anybody wondering what the devices in question are, see: http://www.anandtech.com/storage/showdoc.aspx?i=2480 http://www.gigabyte.com.tw/Products/Storage/Products_Overview.aspx?ProductID=2180 -- Craig Ringer -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
