Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22
This is *very* helpful, thanks for taking the time Larry! Looking forward to giving feedback once we have the cluster up. P On Thu, Dec 17, 2009 at 11:23 AM, Tejas N. Bhise te...@gluster.com wrote: Thanks, Larry, for the comprehensive information. Phil, I hope that answers a lot of your questions. Feel free to ask more, we have a great community here. Regards, Tejas. - Original Message - From: Larry Bates larry.ba...@vitalesafe.com To: gluster-users@gluster.org, p...@cryer.us Sent: Thursday, December 17, 2009 9:47:30 PM GMT +05:30 Chennai, Kolkata, Mumbai, New Delhi Subject: Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22 Phi.l, I think the real question you need to ask has to do with why we are using GlusterFS at all and what happens when something fails. Normally GlusterFS is used to provide scalability, redundancy/recovery, and performance. For many applications performance will be the least of the worries so we concentrate on scalability and redundancy/recovery. Scalability can be achieved no matter which way you configure your servers. Using distribute translator (DHT) you can unify all the servers into a single virtual storage space. The problem comes when you look at what happens when you have a machine/drive failures and need the redundancy/recovery capabilities of GlusterFS. By putting 36Tb of storage on a single server and exposing it as a single volume (using either hardware or software RAID), you will have to replicate that to a replacement server after a failure. Replicating 36Tb will take a lot of time and CPU cycles. If you keep things simple (JBOD) and use AFR to replicate drives between servers and use DHT to unify everything together, now you only have to move 1.5Tb/2Tb when a drive fails. You will also note that you get to use 100% of your disk storage this way instead of wasting 1 drive per array with RAID5 or two drives with RAID6. Normally with RAID5/6 it is also imperative that you have a hot spare per array, which means you waste an additional driver per array. To make RAID5/6 work with no single point of failure you have to do something like RAID50/60 across two controllers which gets expensive and much more difficult to manage and to grow. Implementing GlusterFS using more modest hardware makes all those issues go away. Just use GlusterFS to provide the RAID-like capabilities (via AFR and DHT). Personally I doubt that I would set up my storage the way you describe. I probably would (and have) set it up with more smaller servers. Something like three times as many 2U servers with 8x2Tb drives each (or even 6 times as many 1U servers with 4x2Tb drives each) and forget the expensive RAID SATA controllers, they aren't necessary and are just a single point of failure that you can eliminate. In addition you will enjoy significant performance improvements because you have: 1) Many parallel paths to storage (36x1U or 18x2U vs 6x5U servers). Gigabit Ethernet is fast, but still will limit bandwidth to a single machine. 2) Write performance on RAID5/6 is never going to be as fast as JBOD. 3) You should have much more memory caching available (36x8Gb = 256Gb memory or 18x8Gb memory = 128Gb vs maybe 6x16Gb = 96Gb) 4) Management of the storage is done in one place..GlusterFS. No messy RAID controller setups to document/remember. 5) You can expand in the future in a much more granular and controlled fashion. Add 2 machines (1 for replication) and you get 8Tb (using 2Tb drives) of storage. When you want to replace a machine, just set up new one, fail the old one, and let GlusterFS build the new one for you (AFR will do the heavy lifting). CPUs will get faster, hard drives will get faster and bigger in the future, so make it easy to upgrade. A small number of BIG machines makes it a lot harder to do upgrades as new hardware becomes available. 6) Machine failures (motherboard, power supply, etc.) will effect much less of your storage network. Having a spare 1U machine around as a hot spare doesn't cost much (maybe $1200). Having a spare 5U monster around does (probably close to $6000). IMHO 36 x 1U or 18 x 2U servers shouldn't cost any more (and maybe less) than the big boxes you are looking to buy. They are commodity items. If you go the 1U route you don't need anything but a machine, with memory and 4 hard drives (all server motherboards come with at least 4 SATA ports). By using 2Tb drives, I think you would find that the cost would be actually less. By NOT using hardware RAID you can also NOT use RAID-class hard drives which cost about $100 each more than non-RAID hard drives. Just that change alone could save you 6 x 24 = 144 x $100 = $14,400! JBOD just doesn't need RAID-class hard drives because you don't need the sophisticated firmware that the RAID-class hard drives provide. You still will want quality hard drives, but failures will have such a low impact
Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22
Larry All, I would much rather rebuild a bad drive with a raid controller then have to wait for Gluster to do it. With a large number of files doing a ls -aglR can take weeks. Also you don't NEED enterprise drives with a raid controller, i use desktop 1.5tb Seagate drives which happy as a clam on a 3ware SAS card under a SAS expander. liam On Thu, Dec 17, 2009 at 8:17 AM, Larry Bates larry.ba...@vitalesafe.com wrote: Phi.l, I think the real question you need to ask has to do with why we are using GlusterFS at all and what happens when something fails. Normally GlusterFS is used to provide scalability, redundancy/recovery, and performance. For many applications performance will be the least of the worries so we concentrate on scalability and redundancy/recovery. Scalability can be achieved no matter which way you configure your servers. Using distribute translator (DHT) you can unify all the servers into a single virtual storage space. The problem comes when you look at what happens when you have a machine/drive failures and need the redundancy/recovery capabilities of GlusterFS. By putting 36Tb of storage on a single server and exposing it as a single volume (using either hardware or software RAID), you will have to replicate that to a replacement server after a failure. Replicating 36Tb will take a lot of time and CPU cycles. If you keep things simple (JBOD) and use AFR to replicate drives between servers and use DHT to unify everything together, now you only have to move 1.5Tb/2Tb when a drive fails. You will also note that you get to use 100% of your disk storage this way instead of wasting 1 drive per array with RAID5 or two drives with RAID6. Normally with RAID5/6 it is also imperative that you have a hot spare per array, which means you waste an additional driver per array. To make RAID5/6 work with no single point of failure you have to do something like RAID50/60 across two controllers which gets expensive and much more difficult to manage and to grow. Implementing GlusterFS using more modest hardware makes all those issues go away. Just use GlusterFS to provide the RAID-like capabilities (via AFR and DHT). Personally I doubt that I would set up my storage the way you describe. I probably would (and have) set it up with more smaller servers. Something like three times as many 2U servers with 8x2Tb drives each (or even 6 times as many 1U servers with 4x2Tb drives each) and forget the expensive RAID SATA controllers, they aren't necessary and are just a single point of failure that you can eliminate. In addition you will enjoy significant performance improvements because you have: 1) Many parallel paths to storage (36x1U or 18x2U vs 6x5U servers). Gigabit Ethernet is fast, but still will limit bandwidth to a single machine. 2) Write performance on RAID5/6 is never going to be as fast as JBOD. 3) You should have much more memory caching available (36x8Gb = 256Gb memory or 18x8Gb memory = 128Gb vs maybe 6x16Gb = 96Gb) 4) Management of the storage is done in one place..GlusterFS. No messy RAID controller setups to document/remember. 5) You can expand in the future in a much more granular and controlled fashion. Add 2 machines (1 for replication) and you get 8Tb (using 2Tb drives) of storage. When you want to replace a machine, just set up new one, fail the old one, and let GlusterFS build the new one for you (AFR will do the heavy lifting). CPUs will get faster, hard drives will get faster and bigger in the future, so make it easy to upgrade. A small number of BIG machines makes it a lot harder to do upgrades as new hardware becomes available. 6) Machine failures (motherboard, power supply, etc.) will effect much less of your storage network. Having a spare 1U machine around as a hot spare doesn't cost much (maybe $1200). Having a spare 5U monster around does (probably close to $6000). IMHO 36 x 1U or 18 x 2U servers shouldn't cost any more (and maybe less) than the big boxes you are looking to buy. They are commodity items. If you go the 1U route you don't need anything but a machine, with memory and 4 hard drives (all server motherboards come with at least 4 SATA ports). By using 2Tb drives, I think you would find that the cost would be actually less. By NOT using hardware RAID you can also NOT use RAID-class hard drives which cost about $100 each more than non-RAID hard drives. Just that change alone could save you 6 x 24 = 144 x $100 = $14,400! JBOD just doesn't need RAID-class hard drives because you don't need the sophisticated firmware that the RAID-class hard drives provide. You still will want quality hard drives, but failures will have such a low impact that it is much less of a problem. By using more smaller machines you also eliminate the need for redundant power supplies (which would be a requirement in your large boxes because it would be a single point of failure on a large
Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22
Author is exaggerating. We recover 6 TB RAID-5 on desktop-class hardware in less then 6 hours. Our RAID is controlled by LSR (Linux Software RAID). Performance is not good while rebuilding a single node, but GlusterFS replicate/distribute translators help. Arvids Godjuks wrote: Consider this - a rebuild of 1.5-2 TB HDD in raid5/6 array can easily take up to few days to complete. At that moment your storage at that node will not perform well. I read a week ago very good article with research of this area, only thing it's in russian, but it mentions a few english sources too. Maybe google translate will help. Here's the original link: http://habrahabr.ru/blogs/hardware/78311/ Here's the google translate version: http://translate.google.com/translate?js=yprev=_thl=enie=UTF-8layout=1eotf=1u=http%3A%2F%2Fhabrahabr.ru%2Fblogs%2Fhardware%2F78311%2Fsl=rutl=en (looks quite neet by the way) 2010/1/5 Liam Slusser lslus...@gmail.com: Larry All, I would much rather rebuild a bad drive with a raid controller then have to wait for Gluster to do it. With a large number of files doing a ls -aglR can take weeks. Also you don't NEED enterprise drives with a raid controller, i use desktop 1.5tb Seagate drives which happy as a clam on a 3ware SAS card under a SAS expander. liam ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22
2010/1/6 Liam Slusser lslus...@gmail.com: Arvids Larry, Interesting read, Arvids. ... Well, I think JBOD style makes it much more less painfull loosing a disk than a whole RAID array. Single disk will be restored far faster than a whole raid array. Especially if you can do hotswap of your disks, well yes, hardware should support this ofcourse. Anyway, I do think that a correct combination of AFR and DHT will make up for that disk loss. If only Gluster could do data relocation as a node/disk goes offline. Konstantin Sharlaimov RAID-6 isn't raid-5. Raid-6 has more parity control disks, that means it can do paralel reads from more than one disk and performance doesn't degrade as much as with raid-5. That's why I think you get your rstore fast - your disk is able to recive data fast enought to do writes at high speed: actualy 8 hours for 1.5TB is about 72MB/sec average write speed - not many disks can keep up such speeds all the time. ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22
Yeah im waiting for Gluster to come out with a 3.0.1 release before i upgrade. I'll make sure to do my best to compare 3.0.1 with OneFS's performance/recovery/etc once i upgrade. I still have two Isilon clusters which aren't in production anymore in our lab i can play around with. And i've been waiting for brfs for awhile now, it can't come soon enough! thanks, liam On Tue, Jan 5, 2010 at 7:48 PM, Harshavardhana har...@gluster.com wrote: Hi Liam, GlusterFS does checksum based self-heal since the 3.0 release, i would believe your experiences are from 2.0? which has issues of doing a full file self-heal which will a lot of time. But i would suggest an upgrade with 3.0.1 release which is due Feb 1st week for your cluster. 3.x releases with new self-heal you should get very less rebuild times. If its possible to compare the 3.0.1 rebuild times with the One-FS from Isilon should help us improve it too. Thanks \ I would suggest wait for brtfs. ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22
Thanks, Larry, for the comprehensive information. Phil, I hope that answers a lot of your questions. Feel free to ask more, we have a great community here. Regards, Tejas. - Original Message - From: Larry Bates larry.ba...@vitalesafe.com To: gluster-users@gluster.org, p...@cryer.us Sent: Thursday, December 17, 2009 9:47:30 PM GMT +05:30 Chennai, Kolkata, Mumbai, New Delhi Subject: Re: [Gluster-users] Gluster-users Digest, Vol 20, Issue 22 Phi.l, I think the real question you need to ask has to do with why we are using GlusterFS at all and what happens when something fails. Normally GlusterFS is used to provide scalability, redundancy/recovery, and performance. For many applications performance will be the least of the worries so we concentrate on scalability and redundancy/recovery. Scalability can be achieved no matter which way you configure your servers. Using distribute translator (DHT) you can unify all the servers into a single virtual storage space. The problem comes when you look at what happens when you have a machine/drive failures and need the redundancy/recovery capabilities of GlusterFS. By putting 36Tb of storage on a single server and exposing it as a single volume (using either hardware or software RAID), you will have to replicate that to a replacement server after a failure. Replicating 36Tb will take a lot of time and CPU cycles. If you keep things simple (JBOD) and use AFR to replicate drives between servers and use DHT to unify everything together, now you only have to move 1.5Tb/2Tb when a drive fails. You will also note that you get to use 100% of your disk storage this way instead of wasting 1 drive per array with RAID5 or two drives with RAID6. Normally with RAID5/6 it is also imperative that you have a hot spare per array, which means you waste an additional driver per array. To make RAID5/6 work with no single point of failure you have to do something like RAID50/60 across two controllers which gets expensive and much more difficult to manage and to grow. Implementing GlusterFS using more modest hardware makes all those issues go away. Just use GlusterFS to provide the RAID-like capabilities (via AFR and DHT). Personally I doubt that I would set up my storage the way you describe. I probably would (and have) set it up with more smaller servers. Something like three times as many 2U servers with 8x2Tb drives each (or even 6 times as many 1U servers with 4x2Tb drives each) and forget the expensive RAID SATA controllers, they aren't necessary and are just a single point of failure that you can eliminate. In addition you will enjoy significant performance improvements because you have: 1) Many parallel paths to storage (36x1U or 18x2U vs 6x5U servers). Gigabit Ethernet is fast, but still will limit bandwidth to a single machine. 2) Write performance on RAID5/6 is never going to be as fast as JBOD. 3) You should have much more memory caching available (36x8Gb = 256Gb memory or 18x8Gb memory = 128Gb vs maybe 6x16Gb = 96Gb) 4) Management of the storage is done in one place..GlusterFS. No messy RAID controller setups to document/remember. 5) You can expand in the future in a much more granular and controlled fashion. Add 2 machines (1 for replication) and you get 8Tb (using 2Tb drives) of storage. When you want to replace a machine, just set up new one, fail the old one, and let GlusterFS build the new one for you (AFR will do the heavy lifting). CPUs will get faster, hard drives will get faster and bigger in the future, so make it easy to upgrade. A small number of BIG machines makes it a lot harder to do upgrades as new hardware becomes available. 6) Machine failures (motherboard, power supply, etc.) will effect much less of your storage network. Having a spare 1U machine around as a hot spare doesn't cost much (maybe $1200). Having a spare 5U monster around does (probably close to $6000). IMHO 36 x 1U or 18 x 2U servers shouldn't cost any more (and maybe less) than the big boxes you are looking to buy. They are commodity items. If you go the 1U route you don't need anything but a machine, with memory and 4 hard drives (all server motherboards come with at least 4 SATA ports). By using 2Tb drives, I think you would find that the cost would be actually less. By NOT using hardware RAID you can also NOT use RAID-class hard drives which cost about $100 each more than non-RAID hard drives. Just that change alone could save you 6 x 24 = 144 x $100 = $14,400! JBOD just doesn't need RAID-class hard drives because you don't need the sophisticated firmware that the RAID-class hard drives provide. You still will want quality hard drives, but failures will have such a low impact that it is much less of a problem. By using more smaller machines you also eliminate the need for redundant power supplies (which would be a requirement in your large boxes because it would be a single point