Thread: opinion on disk speed
I have a choice to make on a RAID enclosure: 14x 36GB 15kRPM ultra 320 SCSI drives OR 12x 72GB 10kRPM ultra 320 SCSI drives both would be configured into RAID 10 over two SCSI channels using a megaraid 320-2x card. My goal is speed. Either would provide more disk space than I would need over the next two years. The database does a good number of write transactions, and a decent number of sequential scans over the whole DB (about 60GB including indexes) for large reports. My only concern is the 10kRPM vs 15kRPM. The advantage of the 10k disks is that it would come from the same vendor as the systems to which it will be connected, making procurement easier.
> -----Original Message----- > From: pgsql-performance-owner@postgresql.org > [mailto:pgsql-performance-owner@postgresql.org] On Behalf Of > Vivek Khera > Sent: 08 December 2005 16:52 > To: Postgresql Performance > Subject: [PERFORM] opinion on disk speed > > I have a choice to make on a RAID enclosure: > > 14x 36GB 15kRPM ultra 320 SCSI drives > > OR > > 12x 72GB 10kRPM ultra 320 SCSI drives > > both would be configured into RAID 10 over two SCSI channels using a > megaraid 320-2x card. > > My goal is speed. Either would provide more disk space than I would > need over the next two years. > > The database does a good number of write transactions, and a decent > number of sequential scans over the whole DB (about 60GB including > indexes) for large reports. > > My only concern is the 10kRPM vs 15kRPM. The advantage of the 10k > disks is that it would come from the same vendor as the systems to > which it will be connected, making procurement easier. 15K drives (well, the Seagate Cheetah X15's that I have a lot of at least) can run very hot compared to the 10K's. Might be worth bearing (no pun intended) in mind. Other than that, without knowing the full specs of the drives, you've got 2 extra spindles and a probably-lower-seek time if you go for the X15's so that would seem likely to be the faster option. Regards, Dave
On Thu, 2005-12-08 at 10:52, Vivek Khera wrote: > I have a choice to make on a RAID enclosure: > > 14x 36GB 15kRPM ultra 320 SCSI drives > > OR > > 12x 72GB 10kRPM ultra 320 SCSI drives > > both would be configured into RAID 10 over two SCSI channels using a > megaraid 320-2x card. > > My goal is speed. Either would provide more disk space than I would > need over the next two years. > > The database does a good number of write transactions, and a decent > number of sequential scans over the whole DB (about 60GB including > indexes) for large reports. > > My only concern is the 10kRPM vs 15kRPM. The advantage of the 10k > disks is that it would come from the same vendor as the systems to > which it will be connected, making procurement easier. I would say that the RAID controller and the amount of battery backed cache will have a greater impact than the difference in seek times on those two drives. Also, having two more drives in the 15k category is likely to play to its advantage more so than the speed of the drive spindles and seek times. If you're worried about higher failures due to heat etc... you could always make a couple of the drives spares. Looking at the datasheet for the seagate 10k and 15k drives, it would appear there is another difference, The 10k drives list a sustained xfer rate of 39 to 80 MBytes / second, while the 15k drives list one of 58 to 96. That's quite a bit faster. So, sequential scans should be faster as well. Power consumption isn't much differnt, about a watt more for the 15ks, so that's no big deal. I'd do a bit of googling to see if there are a lot more horror stories with 15k drives than with the 10k ones.
On Thu, 2005-12-08 at 11:52 -0500, Vivek Khera wrote: > I have a choice to make on a RAID enclosure: > > 14x 36GB 15kRPM ultra 320 SCSI drives > > OR > > 12x 72GB 10kRPM ultra 320 SCSI drives > > both would be configured into RAID 10 over two SCSI channels using a > megaraid 320-2x card. > > My goal is speed. Either would provide more disk space than I would > need over the next two years. > > The database does a good number of write transactions, and a decent > number of sequential scans over the whole DB (about 60GB including > indexes) for large reports. The STR of 15k is quite a bit higher than 10k. I'd be inclined toward the 15k if it doesn't impact the budget. For the write transactions, the speed and size of the DIMM on that LSI card will matter the most. I believe the max memory on that adapter is 512MB. These cost so little that it wouldn't make sense to go with anything smaller. When comparing the two disks, don't forget to check for supported SCSI features. In the past I've been surprised that some 10k disks don't support packetization, QAS, and so forth. All 15k disks seem to support these. Don't forget to post some benchmarks when your vendor delivers ;) -jwb
On Thu, 8 Dec 2005, Vivek Khera wrote: > I have a choice to make on a RAID enclosure: > > 14x 36GB 15kRPM ultra 320 SCSI drives > > OR > > 12x 72GB 10kRPM ultra 320 SCSI drives > > both would be configured into RAID 10 over two SCSI channels using a megaraid > 320-2x card. > > My goal is speed. Either would provide more disk space than I would need > over the next two years. > > The database does a good number of write transactions, and a decent number of > sequential scans over the whole DB (about 60GB including indexes) for large > reports. > > My only concern is the 10kRPM vs 15kRPM. The advantage of the 10k disks is > that it would come from the same vendor as the systems to which it will be > connected, making procurement easier. if space isn't an issue then you fall back to the old standby rules of thumb more spindles are better (more disk heads that can move around independantly) faster drives are better (less time to read or write a track) so the 15k drive option is better one other note, you probably don't want to use all the disks in a raid10 array, you probably want to split a pair of them off into a seperate raid1 array and put your WAL on it. David Lang
> one other note, you probably don't want to use all the disks in a raid10 > array, you probably want to split a pair of them off into a seperate > raid1 array and put your WAL on it. Is a RAID 1 array of two disks sufficient for WAL? What's a typical setup for a high performance PostgreSQL installation? RAID 1 for WAL and RAID 10 for data? I've read that splitting the WAL and data offers huge performance benefits. How much additional benefit is gained by moving indexes to another RAID array? Would you typically set the indexes RAID array up as RAID 1 or 10?
On Thu, 8 Dec 2005 17:03:27 -0000 "Dave Page" <dpage@vale-housing.co.uk> wrote: > > -----Original Message----- > > From: pgsql-performance-owner@postgresql.org > > [mailto:pgsql-performance-owner@postgresql.org] On Behalf Of > > Vivek Khera > > > I have a choice to make on a RAID enclosure: > > > > 14x 36GB 15kRPM ultra 320 SCSI drives > > > > OR > > > > 12x 72GB 10kRPM ultra 320 SCSI drives > > > > both would be configured into RAID 10 over two SCSI channels using > > a megaraid 320-2x card. > > 15K drives (well, the Seagate Cheetah X15's that I have a lot of at > least) can run very hot compared to the 10K's. Might be worth bearing > (no pun intended) in mind. > > Other than that, without knowing the full specs of the drives, you've > got 2 extra spindles and a probably-lower-seek time if you go for the > X15's so that would seem likely to be the faster option. I agree, the extra spindles and lower seek times are better if all you are concerned about is raw speed. However, that has to be balanced, from an overall perspective, with the nice single point of ordering/contact/support/warranty of the one vendor. It's a tough call. --------------------------------- Frank Wiles <frank@wiles.org> http://www.wiles.org ---------------------------------
On Thu, 08 Dec 2005 11:50:33 -0600 Scott Marlowe <smarlowe@g2switchworks.com> wrote: > Power consumption isn't much differnt, about a watt more for the 15ks, > so that's no big deal. I'd do a bit of googling to see if there are a > lot more horror stories with 15k drives than with the 10k ones. Just an FYI, but I've run both 10k and 15k rpm drives in PostgreSQL servers and haven't experienced any "horror stories". They do run hotter, but this shouldn't be a big issue in a decent case in a typical server room environment. --------------------------------- Frank Wiles <frank@wiles.org> http://www.wiles.org ---------------------------------
Frank Wiles wrote: > > > I agree, the extra spindles and lower seek times are better if all > you are concerned about is raw speed. > > However, that has to be balanced, from an overall perspective, with > the nice single point of ordering/contact/support/warranty of the > one vendor. It's a tough call. Well, if your favourite dealer can't supply you with such common equipment as 15k drives you should consider changing the dealer. They don't seem to be aware of db hardware reqirements. Regards, Andreas
On Fri, 09 Dec 2005 09:15:25 -0500 "Jeremy Haile" <jhaile@fastmail.fm> wrote: > > one other note, you probably don't want to use all the disks in a > > raid10 array, you probably want to split a pair of them off into a > > seperate raid1 array and put your WAL on it. > > Is a RAID 1 array of two disks sufficient for WAL? What's a typical > setup for a high performance PostgreSQL installation? RAID 1 for WAL > and RAID 10 for data? > > I've read that splitting the WAL and data offers huge performance > benefits. How much additional benefit is gained by moving indexes to > another RAID array? Would you typically set the indexes RAID array up > as RAID 1 or 10? Yes most people put the WAL on a RAID 1 and use all the remaining disks in RAID 10 for data. Whether or not moving your indexes onto a different RAID array is worthwhile is harder to judge. If your indexes are small enough that they will usually be in ram, but your data is to large to fit then having the extra spindles available on the data partition is probably better. As always, it is probably best to test both configurations to see which is optimal for your particular application and setup. --------------------------------- Frank Wiles <frank@wiles.org> http://www.wiles.org ---------------------------------
On Dec 9, 2005, at 10:50 AM, Andreas Pflug wrote: > Well, if your favourite dealer can't supply you with such common > equipment as 15k drives you should consider changing the dealer. > They don't seem to be aware of db hardware reqirements. Thanks to all for your opinions. I'm definitely sticking with 15k drives like I've done in the past for all my other servers. The reason I considered the 10k was because of the simplicity of ordering from the same vendor. They do offer 15k drives, but at double the capacity I needed (73GB drives) which would make the cost way high and overkill for what I need.
On Dec 8, 2005, at 2:21 PM, Jeffrey W. Baker wrote: > For the write transactions, the speed and size of the DIMM on that LSI > card will matter the most. I believe the max memory on that > adapter is > 512MB. These cost so little that it wouldn't make sense to go with > anything smaller. From where did you get LSI MegaRAID controller with 512MB? The 320-2X doesn't seem to come with more than 128 from the factory. Can you just swap out the DIMM card for higher capacity?
On Dec 12, 2005, at 1:59 PM, Vivek Khera wrote: > From where did you get LSI MegaRAID controller with 512MB? The > 320-2X doesn't seem to come with more than 128 from the factory. > > Can you just swap out the DIMM card for higher capacity? We've swapped out the DIMMs on MegaRAID controllers. Given the cost of a standard low-end DIMM these days (which is what the LSI controllers use last I checked), it is a very cheap upgrade. Admittedly I've never actually run benchmarks to see if it made a significant difference in practice, but it certainly seems like it should in theory and the upgrade cost is below the noise floor for most database servers. J. Andrew Rogers
On Dec 12, 2005, at 5:16 PM, J. Andrew Rogers wrote: > We've swapped out the DIMMs on MegaRAID controllers. Given the > cost of a standard low-end DIMM these days (which is what the LSI > controllers use last I checked), it is a very cheap upgrade. What's the max you can put into one of these cards? I haven't been able to find docs on which specific DIMM type they use... Thanks!
On Mon, 2005-12-12 at 16:19, Vivek Khera wrote: > On Dec 12, 2005, at 5:16 PM, J. Andrew Rogers wrote: > > > We've swapped out the DIMMs on MegaRAID controllers. Given the > > cost of a standard low-end DIMM these days (which is what the LSI > > controllers use last I checked), it is a very cheap upgrade. > > What's the max you can put into one of these cards? I haven't been > able to find docs on which specific DIMM type they use... I found the manual for the 4 port U320 SCSI controller, and it listed 256 Meg for single data rate DIMM, and 512 Meg for DDR DIMM. This was on the lsi at: http://www.lsilogic.com/files/docs/techdocs/storage_stand_prod/RAIDpage/mr_320_ug.pdf I believe. They've got a new one coming out, that's SAS, like SCSI on SATA or something. It comes with 256Meg, removeable, but doesn't yet say what the max size it. I'd love to have one of these that could hold a couple of gigs for a TPC type test.
On Dec 12, 2005, at 2:19 PM, Vivek Khera wrote: > On Dec 12, 2005, at 5:16 PM, J. Andrew Rogers wrote: > >> We've swapped out the DIMMs on MegaRAID controllers. Given the >> cost of a standard low-end DIMM these days (which is what the LSI >> controllers use last I checked), it is a very cheap upgrade. > > What's the max you can put into one of these cards? I haven't been > able to find docs on which specific DIMM type they use... Table 3.7 in the MegaRAID Adapter User's Guide has the specs and limits for various controllers. For the 320-2x, the limit is 512MB of PC100 ECC RAM. J. Andrew Rogers