From owner-chemistry@ccl.net Sat Dec 13 02:50:00 2008 From: "Kalju Kahn kalju- -chem.ucsb.edu" To: CCL Subject: CCL: Hard drives and linux Message-Id: <-38293-081213024339-11575-h65nYlogQQrtNsJG9qauQg*server.ccl.net> X-Original-From: "Kalju Kahn" Content-Transfer-Encoding: 8bit Content-Type: text/plain;charset=iso-8859-1 Date: Fri, 12 Dec 2008 23:39:22 -0800 (PST) MIME-Version: 1.0 Sent to CCL by: "Kalju Kahn" [kalju.*o*.chem.ucsb.edu] Hi John, Couple of points: 1) I suggest RAID 0 for your scratch directory. I have 2X 500 GB WD SATA II disks as 1 TB linux softraid and I filled it yesterday with a single MP optimization job ... so do not get much less than this. With RAID 0 and newer disks you would be getting transfer rates at least 160 MB/sec, this is better than any single 10K RPM disk. 2) Do not worry about jumpers and buffer transfer rates of 150 (SATA I) vs 300 (SATA II). Your data transfers to disk surface will be less than 150 MB with current technology; the advertised 300 GB/sec transfer rate to 16 or 32 MB buffer is meaningless when you are writing gigabytes of integrals. But get a SATA II disk simply because this is how new disks are made nowadays. 3) There are two WD 1 TB green drives. The old WD10EACS is lousy, avoid this one. The newer WD10EADS (http://www.tomshardware.com/reviews/green-terabyte-1tb,2078.html) is a good choice if you are worried about power consumption and the machine is idle a long time (e.g. home computer that runs MP2 jobs part time. For 24/7 operation consider a pair of WD1002FBYS (Raid Edition 3, see http://www.tomshardware.com/reviews/hdd-terabyte-1tb,2077-8.html). With RAID 0 you should get hit over 200 GB/sec transfer rates. Hope this helps, Kalju > CClers, > > Hopefully this is not too far out of CCL's area, but then again, many of > have to be [seudo sys mgr at the same time. I am starting to run MP2 > geometry optimizations, and there ia a lot of disk use, and I want to > either > stripe disks or use a 10KRPM scratch drive. The Western Digital "green" > 10KRPM drives have been recommended, and fewer drives are advantageous... > HOWEVER, WD says that they do not support Linux [their drives SATA II > 300GBS > _can_ be used if the drive is jumpered, which knocks it down to a 150GBS > capability. > > Thoughts/experience anyone? > > Many thanks, > > John McKelvey > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Dr. Kalju Kahn Department of Chemistry and Biochemistry UC Santa Barbara, CA 93106 From owner-chemistry@ccl.net Sat Dec 13 15:21:01 2008 From: "Justin Finnerty justin.finnerty() uni-oldenburg.de" To: CCL Subject: CCL:G: Hard drives and linux Message-Id: <-38294-081213151910-31122-g1HPLHiiW5mZlgPUlPutBw[-]server.ccl.net> X-Original-From: "Justin Finnerty" Content-Transfer-Encoding: 8bit Content-Type: text/plain;charset=iso-8859-1 Date: Sat, 13 Dec 2008 21:18:56 +0100 (CET) MIME-Version: 1.0 Sent to CCL by: "Justin Finnerty" [justin.finnerty**uni-oldenburg.de] > CClers, > > Hopefully this is not too far out of CCL's area, but then again, many of > have to be [seudo sys mgr at the same time. I am starting to run MP2 > geometry optimizations, and there ia a lot of disk use, and I want to > either > stripe disks or use a 10KRPM scratch drive. The Western Digital "green" > 10KRPM drives have been recommended, and fewer drives are advantageous... > HOWEVER, WD says that they do not support Linux [their drives SATA II > 300GBS > _can_ be used if the drive is jumpered, which knocks it down to a 150GBS > capability. > > Thoughts/experience anyone? A couple of ideas/experiences. *) First check that your calculation actually is IO bound. Gaussian is pretty good at optimising its disk usage to maximise CPU utilisation. Looking at CPU utilisation during my gaussian MP2 runs a 4-CPU node is >95% utilised with "calculation" with only a single hard-disk. Other codes may offer different performance characteristics, but RAID disks will only offer substantial performance gains if your job is severely IO bound. *) Look at using LVM to create the RAID instead of the kernel raid system. LVM should be just as fast but I find it much easier to work with. *) Look carefully at the settings for the hard-disks, RAID system and filesystem. For a few very large files the larger the stride the better. The last time I looked harddisks could read/write 128k at a time, what they used to call "read-ahead", which would then be the ideal size for a RAID stride. But modern disks may read more, which would make larger strides possible. This then has to be matched with the settings for the filesystem. *) Use the ext2 filesystem. XFS, Reiserfs, ext3 etc are all competing for the fastest data-secure filesystem, but ext2 is the fastest data-insecure filesystem and you don't need data-security for a scratch filesystem. Cheers Justin -- Dr Justin Finnerty Rm W3-1-218 Ph 49 (441) 798 3726 Carl von Ossietzky Universität Oldenburg From owner-chemistry@ccl.net Sat Dec 13 17:02:00 2008 From: "Zoran D Matovic zmatovic-,-kg.ac.rs" To: CCL Subject: CCL:G: Scientific Linux and Gaussian03 Message-Id: <-38295-081213170023-1405-rJ6uWjhxiWd+5G17dcRIog(_)server.ccl.net> X-Original-From: "Zoran D Matovic" Date: Sat, 13 Dec 2008 17:00:19 -0500 Sent to CCL by: "Zoran D Matovic" [zmatovic#%#kg.ac.rs] Hi folks, Does any of you have experience with gaussian03/Linda7.1 running on Scientific linux? I have some problems so if You successifuly ran jobs pls e-mail me. Have a wonderfull day Zoran