Re: [Gluster-users] Interesting experiment
Part of the performance loss is that you cannot get the full 4 Gbit of bandwidth between 2 hosts. Usually you are limited to the throughput of a single link between 2 hosts. And if you are using a round-robin method of bonding, then you run into performance losses due to TCP packets coming in out of order. - Mike -Original Message- From: gluster-users-boun...@gluster.org [mailto:gluster-users-boun...@gluster.org] On Behalf Of Mickey Mazarick Sent: Wednesday, August 19, 2009 4:19 PM To: Nathan Stratton Cc: gluster-users@gluster.org Subject: Re: [Gluster-users] Interesting experiment Just a note we initially tried to set up our storage network with bonded 4 port gig E connections per client and storage node and it was still ~1/3 the speed of infiniband. There also appears to be more overhead in unwrapping data from packets even with jumbo frames set. We did see about a 50% increase in throughput with 2 bonded gig ports, but not double the speed that you would expect. Make sure you use a trunking mechanism and not an active/passive configuration. -Mic Nathan Stratton wrote: On Wed, 19 Aug 2009, Hiren Joshi wrote: Is it worth bonding? This look like I'm maxing out the network connection. Yes, but you should also check out Infiniband. http://www.robotics.net/2009/07/30/infiniband/ Nathan StrattonCTO, BlinkMind, Inc. nathan at robotics.net nathan at blinkmind.com http://www.robotics.nethttp://www.blinkmind.com ___ 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 ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] Interesting experiment
-Original Message- From: Liam Slusser [mailto:lslus...@gmail.com] Sent: 18 August 2009 18:51 To: Hiren Joshi Cc: gluster-users@gluster.org Subject: Re: [Gluster-users] Interesting experiment On Tue, Aug 18, 2009 at 3:05 AM, Hiren Joshij...@moonfruit.com wrote: Hi, Ok, the basic setup is 6 bricks per server, 2 servers. Mirror the six bricks and DHT them. I'm running three tests, dd 1G of zeros to the gluster mount, dd 1000 100k files and dd 1000 1M files. With 3M write-behind I get: 0m35.460s for 1G file 0m52.427s for 100k files 1m37.209s for 1M files Then I added a 400M external journal to all the bricks, the twist being the journals were made on a ram drive Running the same tests: 0m33.614s for 1G file 0m52.851s for 100k files 1m31.693s for 1M files So why is it that adding an external journal (in the ram!) seems to make no difference at all? I would imagine that most of your bottle neck is with the network and not the disks. Modern raid disk storage systems are much quicker than gigabit ethernet. You're right, the raid gives me great (SSD type) performance! This is interesting, I'm on a gigabit network and it looks like it's maxing out when I dd a 1Gig file: about 18 kbits/sec When I dd 1000 1M files: about 8 kbits/sec Is it worth bonding? This look like I'm maxing out the network connection. liam ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] Interesting experiment
Just a note we initially tried to set up our storage network with bonded 4 port gig E connections per client and storage node and it was still ~1/3 the speed of infiniband. There also appears to be more overhead in unwrapping data from packets even with jumbo frames set. We did see about a 50% increase in throughput with 2 bonded gig ports, but not double the speed that you would expect. Make sure you use a trunking mechanism and not an active/passive configuration. -Mic Nathan Stratton wrote: On Wed, 19 Aug 2009, Hiren Joshi wrote: Is it worth bonding? This look like I'm maxing out the network connection. Yes, but you should also check out Infiniband. http://www.robotics.net/2009/07/30/infiniband/ Nathan StrattonCTO, BlinkMind, Inc. nathan at robotics.net nathan at blinkmind.com http://www.robotics.nethttp://www.blinkmind.com ___ 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] Interesting experiment
On Tue, Aug 18, 2009 at 3:05 AM, Hiren Joshij...@moonfruit.com wrote: Hi, Ok, the basic setup is 6 bricks per server, 2 servers. Mirror the six bricks and DHT them. I'm running three tests, dd 1G of zeros to the gluster mount, dd 1000 100k files and dd 1000 1M files. With 3M write-behind I get: 0m35.460s for 1G file 0m52.427s for 100k files 1m37.209s for 1M files Then I added a 400M external journal to all the bricks, the twist being the journals were made on a ram drive Running the same tests: 0m33.614s for 1G file 0m52.851s for 100k files 1m31.693s for 1M files So why is it that adding an external journal (in the ram!) seems to make no difference at all? I would imagine that most of your bottle neck is with the network and not the disks. Modern raid disk storage systems are much quicker than gigabit ethernet. liam ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users