Thread: Consumer-grade vs enterprise-grade disk drives
I'm not sure whether this article has been mentioned here before, but it definitely is worth a read: http://www.usenix.org/events/fast03/tech/full_papers/anderson/anderson_html/ From the proceedings of the FAST '03 conference: More than an interface - SCSI vs. ATADave Anderson, Jim Dykes, Erik RiedelSeagate Research Abstract: This paper sets out to clear up a misconception prominent in the storage community today, that SCSI disc drives and IDE (ATA) disc drives are the same technology internally, and differ only in their external interface and in their suggested retail price. ... This paper attempts to clarify the differences by illuminating some of these design choices and their consequences on final device characteristics. regards, tom lane
Yea, it is a great paper. It is an HTML version of the SCSI PDF I
mentioned a few weeks ago:
http://www.seagate.com/content/docs/pdf/whitepaper/D2c_More_than_Interface_ATA_vs_SCSI_042003.pdf
Someone had mentioned it on the lists a few months ago but I only read
it recently and resuggested it. One problem with the web version is
that the images didn't transfer very clearly. I mention it in my
hardware performance tuning paper.
---------------------------------------------------------------------------
Tom Lane wrote:
> I'm not sure whether this article has been mentioned here before, but
> it definitely is worth a read:
>
> http://www.usenix.org/events/fast03/tech/full_papers/anderson/anderson_html/
>
> >From the proceedings of the FAST '03 conference:
>
> More than an interface - SCSI vs. ATA Dave Anderson, Jim Dykes,
> Erik Riedel Seagate Research
>
> Abstract:
>
> This paper sets out to clear up a misconception prominent in the storage
> community today, that SCSI disc drives and IDE (ATA) disc drives are
> the same technology internally, and differ only in their external
> interface and in their suggested retail price. ... This paper attempts
> to clarify the differences by illuminating some of these design choices
> and their consequences on final device characteristics.
>
> regards, tom lane
>
> ---------------------------(end of broadcast)---------------------------
> TIP 7: don't forget to increase your free space map settings
>
-- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610)
359-1001+ If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square,
Pennsylvania19073
Tom, This is a story that is evolving. Anyone else use StorageReview? Great comprehensive drive benchmarks: http://www.storagereview.com/ Check the comparisons between 15K RPM SCSI drives and the 2004 Western Digital 10K RPM SATA (Raptor) drives. The Raptors are an interesting hybrid of SCSI-related tech and desktop tech, and were some of the first drives with SCSI-like command queuing TCQ/NCQ. I think the last remaining issue in moving to SATA for all enterprise use is the lack of decent SATA controllers, though 3Ware (http://www.3ware.com) is getting there: http://www.3ware.com/link/pdf/Serial-ATA.pdf http://www.3ware.com/products/benchmarks_sata.asp - Luke
Luke Lonergan wrote: > Tom, > > This is a story that is evolving. Anyone else use StorageReview? Great > comprehensive drive benchmarks: > http://www.storagereview.com/ > > Check the comparisons between 15K RPM SCSI drives and the 2004 Western > Digital 10K RPM SATA (Raptor) drives. The Raptors are an interesting hybrid > of SCSI-related tech and desktop tech, and were some of the first drives > with SCSI-like command queuing TCQ/NCQ. > > I think the last remaining issue in moving to SATA for all enterprise use is > the lack of decent SATA controllers, though 3Ware (http://www.3ware.com) is > getting there: > http://www.3ware.com/link/pdf/Serial-ATA.pdf > http://www.3ware.com/products/benchmarks_sata.asp > Although the benchmark numbers are pretty good, they have only published (what looks like) results for sequential IO. It would be interesting to see the random ones, as this would tell us how effective the TCQ implementation is. Cheers Mark
Mark Kirkwood wrote: > Luke Lonergan wrote: > > Although the benchmark numbers are pretty good, they have only published > (what looks like) results for sequential IO. It would be interesting to > see the random ones, as this would tell us how effective the TCQ > implementation is. > Referring to 3ware's benchmark - sorry for the confusion....
On Mon, 2005-05-30 at 21:38 -0700, Luke Lonergan wrote: > Tom, > > This is a story that is evolving. Anyone else use StorageReview? Great > comprehensive drive benchmarks: > http://www.storagereview.com/ > > Check the comparisons between 15K RPM SCSI drives and the 2004 Western > Digital 10K RPM SATA (Raptor) drives. The Raptors are an interesting hybrid > of SCSI-related tech and desktop tech, and were some of the first drives > with SCSI-like command queuing TCQ/NCQ. If we're looking at the same benchmark (File Server DriveMark), the fastest SCSI disk is 65% faster than the fastest SATA disk. The fastest SCSI 10K disk is 25% faster than the SATA. > I think the last remaining issue in moving to SATA for all enterprise use is > the lack of decent SATA controllers, though 3Ware (http://www.3ware.com) is > getting there: > http://www.3ware.com/link/pdf/Serial-ATA.pdf > http://www.3ware.com/products/benchmarks_sata.asp The 3Ware controllers are probably the worst ones you can get for database use. Their caches are slow (I didn't even know that was possible until I bought one), as are the XOR engines. After reading this very comprehensive benchmark: http://print.tweakers.net/?reviews/557 I purchased one of the Areca controllers with a large battery-backed cache. It is unholy fast. I recommend it. -jwb
Jeff, > If we're looking at the same benchmark (File Server DriveMark), the > fastest SCSI disk is 65% faster than the fastest SATA disk. The fastest > SCSI 10K disk is 25% faster than the SATA. I think it's misleading to compare drives on the basis of one benchmark. One of the things I like a lot about storage review is the "Head to Head Comparison" feature. If you use that, you see a completely different story: http://www.storagereview.com/php/benchmark/compare_rtg_2001.php?typeID=10&te stbedID=3&osID=4&raidconfigID=1&numDrives=1&devID_0=237&devID_1=263&devID_2= 259&devCnt=3 In particular, that the 2004 era Raptor is somewhere between the fastest 10K RPM SCSI drive and the fastest (2005) 15K RPM SCSI drive on all but the tests where TCQ is dominant (the two server quite benchmarks). So, if you are doing high concurrent random IOs, you will still prefer the more mature TCQ support in the SCSI drives. If you are doing large block IOs, you will favor the SATA drive for it's faster transfer rate/RPM/$. In the words of the editors in June 2004: " In the end, the potential for SATA to invade the entry- and mid-level server market is there. The performance is definitely there. If the Raptor's reliability proves comparable to the competition and if the infrastructure/support hardware surface, WD will have a viable contender." > The 3Ware controllers are probably the worst ones you can get for > database use. Their caches are slow (I didn't even know that was > possible until I bought one), as are the XOR engines. > > After reading this very comprehensive benchmark: > > http://print.tweakers.net/?reviews/557 Great review - thanks for the link. I'm not seeing as harsh a conclusion about the 3Ware 9500 adapter as you - in fact, it's at the top of the IO/s on IOMeter Fileserver simulation and it's in the pack toward the top on most all of the other benchmarks. The only benchmark it lags significantly on is the RAID5 write test, which correlates with your point about a slow XOR engine. The Areca definitely looks to be the fastest with the highest scaling, but not the fastest at writes either (#3 out of 7, the 3Ware is #4 out of 7). > I purchased one of the Areca controllers with a large battery-backed > cache. It is unholy fast. I recommend it. Cool! Never heard of Areca before this - what did you pay for it? How many drives? What kind of workload do you have? How do you measure performance? - Luke
On T, 2005-05-31 at 17:08 +1200, Mark Kirkwood wrote: > Luke Lonergan wrote: > > Tom, > > > > This is a story that is evolving. Anyone else use StorageReview? Great > > comprehensive drive benchmarks: > > http://www.storagereview.com/ > > > > Check the comparisons between 15K RPM SCSI drives and the 2004 Western > > Digital 10K RPM SATA (Raptor) drives. The Raptors are an interesting hybrid > > of SCSI-related tech and desktop tech, and were some of the first drives > > with SCSI-like command queuing TCQ/NCQ. > > > > I think the last remaining issue in moving to SATA for all enterprise use is > > the lack of decent SATA controllers, though 3Ware (http://www.3ware.com) is > > getting there: > > http://www.3ware.com/link/pdf/Serial-ATA.pdf > > http://www.3ware.com/products/benchmarks_sata.asp > > > > Although the benchmark numbers are pretty good, they have only published > (what looks like) results for sequential IO. It would be interesting to > see the random ones, as this would tell us how effective the TCQ > implementation is. > RAID10 The following are from iozone results for 3Ware with 8 x WD 74G Raptors on 1.6GHz Opteron and 2GB RAM RAID10 "Throughput report Y-axis is type of test X-axis is number of processes" "Record size = 8 Kbytes " "Output is in ops/sec" " Initial write " 1352.90 " Rewrite " 413.31 " Read " 369.01 " Re-read " 368.07 " Reverse Read " 355.94 " Stride read " 358.01 " Random read " 358.29 " Mixed workload " 360.55 " Random write " 359.80 RAID5 "Throughput report Y-axis is type of test X-axis is number of processes" "Record size = 8 Kbytes " "Output is in ops/sec" " Initial write " 1178.55 " Rewrite " 145.91 " Read " 369.06 " Re-read " 364.37 " Reverse Read " 357.83 " Stride read " 357.79 " Random read " 358.04 " Mixed workload " 359.70 " Random write " 360.62 Seems to have either very low overhead for RAID5 or something else is keeping RAID10 speed down. -- Hannu Krosing <hannu@skype.net>
jwbaker@acm.org ("Jeffrey W. Baker") wrote:
> On Mon, 2005-05-30 at 21:38 -0700, Luke Lonergan wrote:
> After reading this very comprehensive benchmark:
>
> http://print.tweakers.net/?reviews/557
>
> I purchased one of the Areca controllers with a large battery-backed
> cache. It is unholy fast. I recommend it.
Do they play well on AMD-64? Good if they do...
--
let name="cbbrowne" and tld="gmail.com" in name ^ "@" ^ tld;;
http://cbbrowne.com/info/slony.html
"I support Microsoft's right to innovate. I just wish they would make
use of that right." - Steve Shaw
Hannu, > RAID10 > > "Throughput report Y-axis is type of test X-axis is number of processes" > "Record size = 8 Kbytes " > "Output is in ops/sec" > " Initial write " 1352.90 > " Rewrite " 413.31 > RAID5 > > "Throughput report Y-axis is type of test X-axis is number of processes" > "Record size = 8 Kbytes " > "Output is in ops/sec" > " Initial write " 1178.55 > " Rewrite " 145.91 > > Seems to have either very low overhead for RAID5 or something else is > keeping RAID10 speed down. The rewrite number is telling - the FPGA on the 3Ware card isn't doing XOR while reading very well. I've found that synthetic benchmarks aren't necessarily predictive for Postgres performance. The issue is that the bottlenecks for writing data and/or scans are not currently in the I/O subsystem. There are many improvements to the executor and to the load/copy and DML paths necessary before this is the case. <begin flame bait :-), answers on bizgres-general> So, I find that for data warehousing workloads (sequential access) if the I/O subsystem runs faster than 100MB/s on sequential writes and about 120MB/s on scans, it will outrun Postgres anyway. We're working to fix this, but right now it appears that high performance I/O on one table (faster than 120MB/s) is not possible with a single instance of Postgres. <end flame bait> Luke