3-4MB/s vinum performance with two Promise PDC20268 UDMA100 controllers
Hi I had four 160G IDE drives in raid5 on a single controller and it worked just fine, never benchmarked it since it was faster than the network anyway. But when adding a second controller card and two more drives the new array has a terrible write performance. I've tried various stripe and block sizes in desperation but that didnt help. Then I've tried assinging the same irq to both controller card in case it was interrupts causing the slow down, so I set both pci slots to use irq3 in the bios (freebsd wants irq 3 for a non existent sio1 port - so I figure that'll be 'free'?) but despite my setting in the bios the cards still show up in dmesg with irq 21 and 22? So I looked thourgh the handbook and tried setting the irq in /boot/device.hints both as hint.atapci.x.irq=3 and hint.ata.x.irq=3 but this didnt work either. The problem is the same in freebsd 5.1 and 5.2 (output below is form 5.2): home# dmesg | grep atapci atapci0: Promise PDC20268 UDMA100 controller port 0xb000-0xb00f,0xb400-0xb403,0xb800-0xb807,0xd000-0xd003,0xd400-0xd407 mem 0xf900-0xf9003fff irq 21 at device 9.0 on pci1 atapci0: [MPSAFE] ata2: at 0xd400 on atapci0 ata3: at 0xb800 on atapci0 atapci1: Promise PDC20268 UDMA100 controller port 0x9400-0x940f,0x9800-0x9803,0xa000-0xa007,0xa400-0xa403,0xa800-0xa807 mem 0xf880-0xf8803fff irq 22 at device 10.0 on pci1 atapci1: [MPSAFE] ata4: at 0xa800 on atapci1 ata5: at 0xa000 on atapci1 atapci2: Intel ICH2 UDMA100 controller port 0x8800-0x880f at device 31.1 on pci0 ata0: at 0x1f0 irq 14 on atapci2 ata1: at 0x170 irq 15 on atapci2 Any thoughts anyone? Bjorn ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]
Re: 3-4MB/s vinum performance with two Promise PDC20268 UDMA100 controllers
Bjorn Eikeland wrote: Hi I had four 160G IDE drives in raid5 on a single controller and it worked just fine, never benchmarked it since it was faster than the network anyway. But when adding a second controller card and two more drives the new array has a terrible write performance. I've tried various stripe and block sizes in desperation but that didnt help. Then I've tried assinging the same irq to both controller card in case it was interrupts causing the slow down, so I set both pci slots to use irq3 in the bios (freebsd wants irq 3 for a non existent sio1 port - so I figure that'll be 'free'?) but despite my setting in the bios the cards still show up in dmesg with irq 21 and 22? So I looked thourgh the handbook and tried setting the irq in /boot/device.hints both as hint.atapci.x.irq=3 and hint.ata.x.irq=3 but this didnt work either. The problem is the same in freebsd 5.1 and 5.2 (output below is form 5.2): home# dmesg | grep atapci atapci0: Promise PDC20268 UDMA100 controller port 0xb000-0xb00f,0xb400-0xb403,0xb800-0xb807,0xd000-0xd003,0xd400-0xd407 mem 0xf900-0xf9003fff irq 21 at device 9.0 on pci1 atapci0: [MPSAFE] ata2: at 0xd400 on atapci0 ata3: at 0xb800 on atapci0 atapci1: Promise PDC20268 UDMA100 controller port 0x9400-0x940f,0x9800-0x9803,0xa000-0xa007,0xa400-0xa403,0xa800-0xa807 mem 0xf880-0xf8803fff irq 22 at device 10.0 on pci1 atapci1: [MPSAFE] ata4: at 0xa800 on atapci1 ata5: at 0xa000 on atapci1 atapci2: Intel ICH2 UDMA100 controller port 0x8800-0x880f at device 31.1 on pci0 ata0: at 0x1f0 irq 14 on atapci2 ata1: at 0x170 irq 15 on atapci2 Any thoughts anyone? Bjorn ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED] Don't assign the same irq to 2 devices, thats a conflict! You may not have used win95, but you don't won't to do that. I know raid 5 isslower than raid 1, but I don't remember any numbers. Also the more complex you make the system the slower it will go, hince raid 5 slower than raid 1. Also pci is a shared bus meaning 1 device talks at a time, so maybe, just maybe if you chipset has a pci bridge because you have like 8 slots, or the make was real kind, you could try card 1 in say slot 2, and card 2 in slot 6? You maybe able to get simultainsuos writes and reads that way. The best option is pcix, or a controler that support 8 drives on its own. Also try to adjust pci latincy or waiting, do a google on it, I hear 95-128 clicks is good. I just noticed the buillt in ide is on pci0 while the rest are on pci1, theres a bridge you can use. If you can get it setup, have 2 drives connected to the onboard ide and the rest to your card. Jason ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]
Re: 3-4MB/s vinum performance with two Promise PDC20268 UDMA100 controllers
On Thursday, 15 January 2004 at 10:44:20 +0100, Bjorn Eikeland wrote: Hi I had four 160G IDE drives in raid5 on a single controller and it worked just fine, never benchmarked it since it was faster than the network anyway. But when adding a second controller card and two more drives the new array has a terrible write performance. ... Any thoughts anyone? How about some information about your configuration and how you measured the performance? Greg -- When replying to this message, please copy the original recipients. If you don't, I may ignore the reply or reply to the original recipients. For more information, see http://www.lemis.com/questions.html See complete headers for address and phone numbers. pgp0.pgp Description: PGP signature
Re: 3-4MB/s vinum performance with two Promise PDC20268 UDMA100 controllers
På Thu, 15 Jan 2004 09:43:01 -0200, skrev jason [EMAIL PROTECTED]: Bjorn Eikeland wrote: Hi I had four 160G IDE drives in raid5 on a single controller and it worked just fine, never benchmarked it since it was faster than the network anyway. But when adding a second controller card and two more drives the new array has a terrible write performance. I've tried various stripe and block sizes in desperation but that didnt help. Then I've tried assinging the same irq to both controller card in case it was interrupts causing the slow down, so I set both pci slots to use irq3 in the bios (freebsd wants irq 3 for a non existent sio1 port - so I figure that'll be 'free'?) but despite my setting in the bios the cards still show up in dmesg with irq 21 and 22? So I looked thourgh the handbook and tried setting the irq in /boot/device.hints both as hint.atapci.x.irq=3 and hint.ata.x.irq=3 but this didnt work either. The problem is the same in freebsd 5.1 and 5.2 (output below is form 5.2): home# dmesg | grep atapci atapci0: Promise PDC20268 UDMA100 controller port 0xb000-0xb00f,0xb400-0xb403,0xb800-0xb807,0xd000-0xd003,0xd400-0xd407 mem 0xf900-0xf9003fff irq 21 at device 9.0 on pci1 atapci0: [MPSAFE] ata2: at 0xd400 on atapci0 ata3: at 0xb800 on atapci0 atapci1: Promise PDC20268 UDMA100 controller port 0x9400-0x940f,0x9800-0x9803,0xa000-0xa007,0xa400-0xa403,0xa800-0xa807 mem 0xf880-0xf8803fff irq 22 at device 10.0 on pci1 atapci1: [MPSAFE] ata4: at 0xa800 on atapci1 ata5: at 0xa000 on atapci1 atapci2: Intel ICH2 UDMA100 controller port 0x8800-0x880f at device 31.1 on pci0 ata0: at 0x1f0 irq 14 on atapci2 ata1: at 0x170 irq 15 on atapci2 Any thoughts anyone? Bjorn ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED] Don't assign the same irq to 2 devices, thats a conflict! You may not have used win95, but you don't won't to do that. I know raid 5 isslower than raid 1, but I don't remember any numbers. Also the more complex you make the system the slower it will go, hince raid 5 slower than raid 1. Also pci is a shared bus meaning 1 device talks at a time, so maybe, just maybe if you chipset has a pci bridge because you have like 8 slots, or the make was real kind, you could try card 1 in say slot 2, and card 2 in slot 6? You maybe able to get simultainsuos writes and reads that way. The best option is pcix, or a controler that support 8 drives on its own. Also try to adjust pci latincy or waiting, do a google on it, I hear 95-128 clicks is good. I just noticed the buillt in ide is on pci0 while the rest are on pci1, theres a bridge you can use. If you can get it setup, have 2 drives connected to the onboard ide and the rest to your card. Jason Thank you for your thoughts Jason! (Your penny is in the mail ;) About the irq thing I think I read (while reading up on bridges) that interrupts were level triggered (as apposed to edge triggered) and thus two NICs sharing a interrupt and asserting it at the same time would only cause one context switch - so I figured it worth a try. As for the pci bus etc, all three pci slots are pci1 - and the secondary onboard ide channel gives me read and write errors (the same type as for a bad udma100 cable - but im sure its the controller as the cable and drives work fine elsewhere). I've found some info on pci latency but will try it tomorrow as the box is headless - just made some refrence measurements tonight. But I acutally remembered that the previos setup wasn't raid5 - the four first drives were striped since it was a temporary arrangement while waiting for the 2nd controller - however I do have a linux machine at home with 4 of the same drives (all on the onboard UDMA100 controller) running raid5 and it does perform better (cant do any mesurements now, but it does accept data at about 60Mbps over a smb share and i think the client maxed out at that). I've done some measurements on the drives with different setups - the results were quite long so I've posted it on a web page instad http://www.eikeland.info/bjorn/archive/040117vinumperf1.txt (Test was dd count=1 bs=65536 if=/dev/zero of=/dev/vinum/test) Briefly summarized: any raid5 (with 3, 4 or 6 drives) writes ~4M/s and reads 27M, 29M and 32M/s respectly. A single drive reads and writes ~40M/s. Raid0 (4 and 6 drives) writes ~50M and reads 50M and 62M/s respectively. Test done to from /dev/zero with 625M or 512M test file. Is this really what write performance I can expect from a raid5 array usch as this? I knew it wouldnt be writing way fast, but I was quite sure it would accept what the 100Mbit network had to offer? (Should I maybe move this over to freebsd-performance?) -Bjorn ___ [EMAIL PROTECTED] mailing list
Vinum performance
Hi all, I was wondering if anyone can tell me about their experience of Vinum and its performance on RAID5 systems? I've built a vinum-raid of 4 160Gb disks that each do a dd if=/dev/zero of=/dev/adNd of about 60Mb/sec. However, when I've set up my raid, performance drops to 6Mb/sec, and with only three disks it drops to 4.5Mb/sec. Is this usual? Because with this speed I cannot run the system at all. Any good way of tuning the system? Cheers Nik ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]
Re: Vinum performance
On Tuesday, 19 August 2003 at 9:56:13 +0200, Niklas Saers Mailinglistaccount wrote: Hi all, I was wondering if anyone can tell me about their experience of Vinum and its performance on RAID5 systems? I've built a vinum-raid of 4 160Gb disks that each do a dd if=/dev/zero of=/dev/adNd of about 60Mb/sec. However, when I've set up my raid, performance drops to 6Mb/sec, and with only three disks it drops to 4.5Mb/sec. Is this usual? No. You should expect about ¼ the write performance of a single disk under these rather contrived circumstances. Because with this speed I cannot run the system at all. Have you built this system for writing single sectors to a volume using dd? Any good way of tuning the system? There are some ideas in the man page and the web site. But in general, RAID-5 isn't designed for heavy write accesses. Greg -- When replying to this message, please copy the original recipients. If you don't, I may ignore the reply or reply to the original recipients. For more information, see http://www.lemis.com/questions.html See complete headers for address and phone numbers pgp0.pgp Description: PGP signature
Re: Vinum performance
Hmmm, I thought the consensus was to use a weird stripe size to avoid getting all the inode/superblock stuff on 1 disk. I seem to recall somebody saying somethingabout using stripe sizes like 273k and such... On Fri, May 30, 2003 at 07:48:45PM +0930, Greg 'groggy' Lehey wrote: On Friday, 30 May 2003 at 11:16:06 +0200, Per olof Ljungmark wrote: Using Vinum on 4.7-RELEASE-p10, I wonder what can be done to optimize performance. So far I am not impressed but perhaps I did not configure Vinum optimal, grateful for any hints thanks. You haven't said what your problem is. It definitely depends on your application (which may simply be the way you measure it). FWIW, the stripe size should be a multiple of the file system block size. Yes, the man pages don't necessarily say that, but it's also not so important. Greg -- When replying to this message, please copy the original recipients. If you don't, I may ignore the reply or reply to the original recipients. For more information, see http://www.lemis.com/questions.html See complete headers for address and phone numbers ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]
Re: Vinum performance
On Monday, 2 June 2003 at 17:48:23 -0700, Jaye Mathisen wrote: On Fri, May 30, 2003 at 07:48:45PM +0930, Greg 'groggy' Lehey wrote: On Friday, 30 May 2003 at 11:16:06 +0200, Per olof Ljungmark wrote: Using Vinum on 4.7-RELEASE-p10, I wonder what can be done to optimize performance. So far I am not impressed but perhaps I did not configure Vinum optimal, grateful for any hints thanks. You haven't said what your problem is. It definitely depends on your application (which may simply be the way you measure it). FWIW, the stripe size should be a multiple of the file system block size. Yes, the man pages don't necessarily say that, but it's also not so important. Hmmm, I thought the consensus was to use a weird stripe size to avoid getting all the inode/superblock stuff on 1 disk. Correct, for some definition of weird. I seem to recall somebody saying somethingabout using stripe sizes like 273k and such... Yes, I once said that. Then it occurred to me that many transfers are complete file system blocks. If you have a stripe size which isn't a multiple of the block size, you'll end up with more transfers split across two devices, which has a negative effect on performance. Greg -- When replying to this message, please copy the original recipients. If you don't, I may ignore the reply or reply to the original recipients. For more information, see http://www.lemis.com/questions.html See complete headers for address and phone numbers pgp0.pgp Description: PGP signature
Re: Vinum performance
Per olof Ljungmark wrote: [ ... ] The bottleneck is write performance, reading from the documentation this is normally a weak point in Vinum/raid5 setups. It's not vinum; RAID-5 write performance is going to be relatively slow even with hardware parity/XOR support. The volume will be used for storing temporary files with sizes in the 1-4MB range, a few hundreds at a time. Reads are fine but writes a bit slow. What ratio of reads to writes do you expect? Copying 170 files total 319MB TO the Vinum volume takes about 2m45s. Copying same files FROM the Vinum volume to another volume on a hardware raid5 controller takes 43s. A 3-1 ratio in speeds for software RAID-5 versus hardware doesn't strike me as being very wrong This machine is not in production yet so I can still make changes to the configuration. The current block size is 16384 and the stripe size 419k. I assume it would be a good idea to change that to for example 491,520? ...or try a few stripe sizes, such as 64K or 128K. -Chuck ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]
Re: Vinum performance
What ratio of reads to writes do you expect? Had no expectations really, I'm new to Vinum. Copying 170 files total 319MB TO the Vinum volume takes about 2m45s. Copying same files FROM the Vinum volume to another volume on a hardware raid5 controller takes 43s. A 3-1 ratio in speeds for software RAID-5 versus hardware doesn't strike me as being very wrong This machine is not in production yet so I can still make changes to the configuration. The current block size is 16384 and the stripe size 419k. I assume it would be a good idea to change that to for example 491,520? ...or try a few stripe sizes, such as 64K or 128K. Chuck, Thanks for your comments, I felt it was a good idea just to check before we start to use this piece seriously then I will not be able to touch it for changing configuration. Will try different stripe sizes as per your advice. /per olof ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]
Re: Vinum performance
- Original Message - From: Per olof Ljungmark [EMAIL PROTECTED] To: Chuck Swiger [EMAIL PROTECTED] Cc: [EMAIL PROTECTED] Sent: Friday, May 30, 2003 8:41 AM What ratio of reads to writes do you expect? Had no expectations really, I'm new to Vinum. Copying 170 files total 319MB TO the Vinum volume takes about 2m45s. Copying same files FROM the Vinum volume to another volume on a hardware raid5 controller takes 43s. A 3-1 ratio in speeds for software RAID-5 versus hardware doesn't strike me as being very wrong This machine is not in production yet so I can still make changes to the configuration. The current block size is 16384 and the stripe size 419k. I assume it would be a good idea to change that to for example 491,520? ...or try a few stripe sizes, such as 64K or 128K. FWIW - The man page suggests avoiding powers of 2 and a minimum 128K strip size. Cheers, Drew Thanks for your comments, I felt it was a good idea just to check before we start to use this piece seriously then I will not be able to touch it for changing configuration. Will try different stripe sizes as per your advice. /per olof ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED] ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]
Re: Vinum performance
On Friday, 30 May 2003 at 11:16:06 +0200, Per olof Ljungmark wrote: Using Vinum on 4.7-RELEASE-p10, I wonder what can be done to optimize performance. So far I am not impressed but perhaps I did not configure Vinum optimal, grateful for any hints thanks. You haven't said what your problem is. It definitely depends on your application (which may simply be the way you measure it). FWIW, the stripe size should be a multiple of the file system block size. Yes, the man pages don't necessarily say that, but it's also not so important. Greg -- When replying to this message, please copy the original recipients. If you don't, I may ignore the reply or reply to the original recipients. For more information, see http://www.lemis.com/questions.html See complete headers for address and phone numbers pgp0.pgp Description: PGP signature
Re: Vinum performance
Greg 'groggy' Lehey wrote: On Friday, 30 May 2003 at 11:16:06 +0200, Per olof Ljungmark wrote: Using Vinum on 4.7-RELEASE-p10, I wonder what can be done to optimize performance. So far I am not impressed but perhaps I did not configure Vinum optimal, grateful for any hints thanks. You haven't said what your problem is. It definitely depends on your application (which may simply be the way you measure it). FWIW, the stripe size should be a multiple of the file system block size. Yes, the man pages don't necessarily say that, but it's also not so important. Sorry for the minimal info given in initial post - The bottleneck is write performance, reading from the documentation this is normally a weak point in Vinum/raid5 setups. The volume will be used for storing temporary files with sizes in the 1-4MB range, a few hundreds at a time. Reads are fine but writes a bit slow. Copying 170 files total 319MB TO the Vinum volume takes about 2m45s. Copying same files FROM the Vinum volume to another volume on a hardware raid5 controller takes 43s. This machine is not in production yet so I can still make changes to the configuration. The current block size is 16384 and the stripe size 419k. I assume it would be a good idea to change that to for example 491,520? If you could please confirm if I could look for further improvements or if the performance is about what one should expect. Many thanks, Per olof ___ [EMAIL PROTECTED] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to [EMAIL PROTECTED]