Re: striping, etc.
On Fri, Jun 20, 1997 at 12:12:14PM -0500, Nathan E Norman wrote: On that linux-raid list I told you about, someone was discussing IDE performance. Seems that with their testing, which may or may not have been very accurate, that putting IDE disks on the same or seperate controllers seemed to have very little difference in performance. I suspect this has more to do with the crappiness of IDE than anything to do with the md algorithms. In fact, depending on your hardware, the performance might even be worse. I have a Shuttle HOT-553, which is a Triton HX motherboard. I have a Western Dig 1.6gb on the primary, and a Quantum 3.2gb which was on the secondary. Linux would initially enable the DMA for the Quantum, then disable it when it mounted the partitions, after a timeout. This doesn't happen now that I've moved the Quantum onto the primary with the WD (although it still happens to my CD-ROM drive which is on the secondary). I don't know if there's a performance difference, but I couldn't be any worse off. So in summary, DMA does not seem to work to drives on the secondary controller, while it works fine on the primary. Hamish -- Hamish Moffatt, StudIEAust[EMAIL PROTECTED] Student, computer science computer systems engineering.3rd year, RMIT. http://hamish.home.ml.org/ (PGP key here) CPOM: [ ] 48% The opposite of a profound truth may well be another profound truth. --Bohr -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
They will go on a machine with 3 200m ide drives, which will be a poor-man's server. My current thinking is to mount / on the first controller, and use the other pair as /usr on the second interface. /usr will be NFS exported. Or would I be better off putting the two /usr drives on separate controllers? I'd think it was better to mount them across separate controllers. With seperate control and data lines, the kernel can issue two simultaneous requests and get data from both at the same time. My understanding with IDE (and EIDE) is that a single controller can only access a single drive at a time and must wait for that request to finish before issuing another. SCSI is a more sophisticated in that it allows a request to be issued and then the bus to idle (for more requests or other data) until the drive finishes processing the request and can blast back the data. This is why SCSI is much better than EIDE when dealing with more than one drive. (At least, this is my understanding... Somebody please correct me if I'm wrong.) Brian ( [EMAIL PROTECTED] ) --- Debian GNU/Linux! Search it at http://insite.verisim.com/search/debian/simple -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
They will go on a machine with 3 200m ide drives, which will be a poor-man's server. My current thinking is to mount / on the first controller, and use the other pair as /usr on the second interface. /usr will be NFS exported. Or would I be better off putting the two /usr drives on separate controllers? I'd think it was better to mount them across separate controllers. With seperate control and data lines, the kernel can issue two simultaneous requests and get data from both at the same time. My understanding with IDE (and EIDE) is that a single controller can only access a single drive at a time and must wait for that request to finish before issuing another. yes; that's the hitch with ide. On the other hand, we don't have spare scsis lying around :) The reason i'm hesitating to put them on separate controllers is that / is also on the first controller. Everything that gets nfs exported will come off /usr, and my concern is that massive hits to the portion that was slaved could leave / unaccesable to the host. SCSI is a more sophisticated in that it allows a request to be issued and then the bus to idle (for more requests or other data) until the drive finishes processing the request and can blast back the data. This is why SCSI is much better than EIDE when dealing with more than one drive. (At least, this is my understanding... Somebody please correct me if I'm wrong.) yes; exactly. I just wish we had scsis. Of coure, if this whole thing works, we may be able to get one . . . rick -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
On Fri, 20 Jun 1997, Rick Hawkins wrote: : : They will go on a machine with 3 200m ide drives, which will be a poor-man's : server. My current thinking is to mount / on the first controller, and : use the other pair as /usr on the second interface. /usr will be NFS : exported. Or would I be better off putting the two /usr drives on : separate controllers? : : I'd think it was better to mount them across separate controllers. With : seperate control and data lines, the kernel can issue two simultaneous : requests and get data from both at the same time. My understanding with : IDE (and EIDE) is that a single controller can only access a single : drive at a time and must wait for that request to finish before issuing : another. : :yes; that's the hitch with ide. On the other hand, we don't have spare :scsis lying around :) : :The reason i'm hesitating to put them on separate controllers is that / :is also on the first controller. Everything that gets nfs exported will :come off /usr, and my concern is that massive hits to the portion that :was slaved could leave / unaccesable to the host. : : : SCSI is a more sophisticated in that it allows a request to be issued : and then the bus to idle (for more requests or other data) until the : drive finishes processing the request and can blast back the data. : : This is why SCSI is much better than EIDE when dealing with more than one : drive. (At least, this is my understanding... Somebody please correct : me if I'm wrong.) : :yes; exactly. I just wish we had scsis. Of coure, if this whole thing :works, we may be able to get one . . . : On that linux-raid list I told you about, someone was discussing IDE performance. Seems that with their testing, which may or may not have been very accurate, that putting IDE disks on the same or seperate controllers seemed to have very little difference in performance. I suspect this has more to do with the crappiness of IDE than anything to do with the md algorithms. Given your concerns about / being accessable, I believe the best choice would be to put both drives on the secondary controller. After all, this is a proof of concept type install, right? I suppose you could try creating a linear device and a raid0 device and run some adhoc tests to see if there's a difference ... but I think you'll find that IDE is holding you back, not the md stuff. Good luck! May the gods grant you many gigs of SCSI disk :) -- Nathan Norman:Hostmaster CFNI:[EMAIL PROTECTED] finger [EMAIL PROTECTED] for PGP public key and other stuff Key fingerprint = CE 03 10 AF 32 81 18 58 9D 32 C2 AB 93 6D C4 72 -- -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
On Fri, 20 Jun 1997, Rick Hawkins wrote: They will go on a machine with 3 200m ide drives, which will be a poor-man's server. My current thinking is to mount / on the first controller, and use the other pair as /usr on the second interface. /usr will be NFS exported. Or would I be better off putting the two /usr drives on separate controllers? I'd think it was better to mount them across separate controllers. With seperate control and data lines, the kernel can issue two simultaneous requests and get data from both at the same time. My understanding with IDE (and EIDE) is that a single controller can only access a single drive at a time and must wait for that request to finish before issuing another. The reason i'm hesitating to put them on separate controllers is that / is also on the first controller. Everything that gets nfs exported will come off /usr, and my concern is that massive hits to the portion that was slaved could leave / unaccesable to the host. Assuming you don't plan to add an ISA IDE/EIDE controller (I've seen them, and I think someone is running Debian with one (giving them three or even four IDE controllers)), I would suggest running both disks on the second controller and using linear, although that's merely a guess, not a result of actual testing. I had a 2x3.1GB RAID0 md array, with both disks as slaves. I had 2x120M swap areas, one on each of the masters. If mirror was running, it was trying to access the (slave) array while swapping to the (master) swap area(s). Horrible performance! Don't think about putting / on md...without a non-md partition, you can't read the kernel, and without the kernel you can't load the md stuff. _DO_ compile md into the kernel, it'll be much easier to use than if it is modularized. One bad note to put forth: Concurrent with a local storm and power failure (though I don't think it was related), my Linux host choked. Upon reboot, fsck on the md array failed (some sort of internal error). With my limited knowledge of filesystems, I couldn't fix it and was forced to rebuild my data. As a result, I chose to remove the md array and downgrade my disk usage (I had a mirror on it, so I just had to give up breathing room and Debian 1.1, only weeks before the release of 1.3 so I wasn't too bummed). In doing so, I rearranged partitions so that / and /usr were on different controllers, swap was on the same disk as /, and got much better performance than before. Very well worth it, but too many changes (and too many other space commitments) to possibly restore the md array and see how performance was after the repartitioning. Sorry I can't verify that speed. And, make sure your drives DON'T spin down ever (I used hdparm -S 0 I think to stop that behavior). HTH, Pete -- Peter J. Templin, Jr. Client Services Analyst Computer Communication Services tel: (717) 524-1590 Bucknell University [EMAIL PROTECTED] -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
Rick Hawkins wrote: wow, that was fast :) I've downloaded it, and read the docs. I compiled the kernel with support for these devices. They will go on a machine with 3 200m ide drives, which will be a poor-man's server. My current thinking is to mount / on the first controller, and use the other pair as /usr on the second interface. /usr will be NFS exported. Or would I be better off putting the two /usr drives on separate controllers? also, would I be better off combining all three? and finally, will linear or raid0 give be better performance on ide drives (i know the answer is raid0 on scsi). Put RAIDed disks on one controller and / and swap on another. Also, IIRC if disks on one controller support different modes, contoller will use the [s]lowest for both. So, if your disks aren't identical you might want to run a few tests on them (hdparm) first. Dimitri -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
Dima [EMAIL PROTECTED] writes: Put RAIDed disks on one controller and / and swap on another. Well, as I understand it, if you're not using hardware raid, specifically, if you're using IDE, then having the RAIDed disks on the same IDE controller mostly defeats the purpose of RAID (at least RAID for acceleration) because IDE can only issue commands to one drive on a given controller at a time. If you're using SCSI, then I'm not sure if it makes a lot of difference if the RAIDed drives are on the same controller or different ones. Probably depends on the quality of the controllers. This is all assuming *software* raid (i.e. md). -- Rob -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
striping, etc.
I've noticed that linux supports volumes across physical devices. However, I haven't figured out which command to use to set this up. I would like to mount a pair of hard drives on the second controller jointly as /usr. THis volume will also be served out by nfs. Could someone give me a hint as to what the command i'm looking for is? and which how-to or faq to look at? rick -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
You want mdtools (the package) and md device support in the kernel, wither compiled in or as a module. Either works well. I have two RAID0 partitions spanned across two 4 GB SCSI drives. Works great. The md commands actually have useful man pages. If you need more info, feel free to email. -- Nathan Norman:Hostmaster CFNI:[EMAIL PROTECTED] finger [EMAIL PROTECTED] for PGP public key and other stuff Key fingerprint = CE 03 10 AF 32 81 18 58 9D 32 C2 AB 93 6D C4 72 -- On Thu, 19 Jun 1997, Rick Hawkins wrote: : :I've noticed that linux supports volumes across physical devices. :However, I haven't figured out which command to use to set this up. I :would like to mount a pair of hard drives on the second controller :jointly as /usr. THis volume will also be served out by nfs. Could :someone give me a hint as to what the command i'm looking for is? and :which how-to or faq to look at? : :rick : : :-- :TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to :[EMAIL PROTECTED] . :Trouble? e-mail to [EMAIL PROTECTED] . : -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
wow, that was fast :) I've downloaded it, and read the docs. I compiled the kernel with support for these devices. They will go on a machine with 3 200m ide drives, which will be a poor-man's server. My current thinking is to mount / on the first controller, and use the other pair as /usr on the second interface. /usr will be NFS exported. Or would I be better off putting the two /usr drives on separate controllers? also, would I be better off combining all three? and finally, will linear or raid0 give be better performance on ide drives (i know the answer is raid0 on scsi). rick -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
Re: striping, etc.
On Thu, 19 Jun 1997, Rick Hawkins wrote: : :wow, that was fast :) : :I've downloaded it, and read the docs. I compiled the kernel with :support for these devices. : :They will go on a machine with 3 200m ide drives, which will be a poor-man's :server. My current thinking is to mount / on the first controller, and :use the other pair as /usr on the second interface. /usr will be NFS :exported. Or would I be better off putting the two /usr drives on :separate controllers? : :also, would I be better off combining all three? and finally, will :linear or raid0 give be better performance on ide drives (i know the :answer is raid0 on scsi). : :rick : My feeling is that you'd do better with seperate controllers - however, I've never run the md devices on IDE (I have an aversion to IDE). I suspect that RAID0 would be faster in any case since it divides access between the drives at all times. There is a mailing list for Linux-raid users. Send mail to [EMAIL PROTECTED], with a body of subscribe linux-raid. It's pretty low volume, and you can find people who have done what you're trying to do. Sorry I can't provide more specific answers, but I avoid IDE like the plague. I'm in the lucky position of having lots of SCSI drives to play with ... I know that's not the norm. Good luck! Linux raid is cool! -- Nathan Norman:Hostmaster CFNI:[EMAIL PROTECTED] finger [EMAIL PROTECTED] for PGP public key and other stuff Key fingerprint = CE 03 10 AF 32 81 18 58 9D 32 C2 AB 93 6D C4 72 -- -- TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word unsubscribe to [EMAIL PROTECTED] . Trouble? e-mail to [EMAIL PROTECTED] .
