Re: [Samba] performance regression between 30.14a and 3.0.20
On Wed, Apr 26, 2006 at 01:20:23PM -0400, Greg Dickie wrote: OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? OK, I just confirmed that iometer opens the files it tests with FILE_WRITE_THROUGH set on the NTCreateX call. That would do it The fix now in SAMBA_3_0 should fix this problem. Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Thu, Apr 27, 2006 at 11:53:29AM +1000, James Peach wrote: On 4/27/06, Greg Dickie [EMAIL PROTECTED] wrote: OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? O_SYNC is not a good option to use if you want good write performance on XFS. It's what you use to when you really want your data safe and don't care about throughput. Ok - it's set in smbd when the client sets the create options flag FILE_WRITE_THROUGH. That was added in the conversion to ntcreateX code from the old openX code. I expected FILE_WRITE_THROUGH to be *very* rare - I'll look over lots of traces to check. I can easily add a check for strict sync = yes before setting that flag. Give me a minute or so Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Wed, Apr 26, 2006 at 01:20:23PM -0400, Greg Dickie wrote: OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? Greg - I LOVE YOU :-). That's almost certainly the problem. Back in the Win9x protocols days (the openX style of calls) we had the problem that Windows clients were setting the sync this data flag for almost every open - completely unneccessary. For 3.0.20x we changed to the NTcreateX style of open, and mapped the create option FILE_WRITE_THROUGH to the O_SYNC. I bet the Windows redirector in W2K, XP and W2K3 is just as dim, and is adding that on almost every open (I'm on a plane right now so can't get to my vmware sessions easily to check). We have an option strict sync which if unset allows us to ignore these calls - I've just added it into the ntcreate open path so we don't set O_SYNC unless strict sync is set. Please check out SAMBA_3_0 and test - this will be in the next 3.0.23 pre-release. Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Actually it was my much smarter colleagues that found this, I'm just the mouth ;-) Thanks for looking at this, I'll check out SAMBA_3_0 when I get back! Greg Jeremy Allison wrote: On Wed, Apr 26, 2006 at 01:20:23PM -0400, Greg Dickie wrote: OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? Greg - I LOVE YOU :-). That's almost certainly the problem. Back in the Win9x protocols days (the openX style of calls) we had the problem that Windows clients were setting the sync this data flag for almost every open - completely unneccessary. For 3.0.20x we changed to the NTcreateX style of open, and mapped the create option FILE_WRITE_THROUGH to the O_SYNC. I bet the Windows redirector in W2K, XP and W2K3 is just as dim, and is adding that on almost every open (I'm on a plane right now so can't get to my vmware sessions easily to check). We have an option strict sync which if unset allows us to ignore these calls - I've just added it into the ntcreate open path so we don't set O_SYNC unless strict sync is set. Please check out SAMBA_3_0 and test - this will be in the next 3.0.23 pre-release. Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On 4/27/06, Jeremy Allison [EMAIL PROTECTED] wrote: On Wed, Apr 26, 2006 at 01:20:23PM -0400, Greg Dickie wrote: OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? Greg - I LOVE YOU :-). That's almost certainly the problem. Back in the Win9x protocols days (the openX style of calls) we had the problem that Windows clients were setting the sync this data flag for almost every open - completely unneccessary. For 3.0.20x we changed to the NTcreateX style of open, and mapped the create option FILE_WRITE_THROUGH to the O_SYNC. I bet the Windows redirector in W2K, XP and W2K3 is just as dim, and is adding that on almost every open (I'm on a plane right now so can't get to my vmware sessions easily to check). xdd explicitly sets this flag when it calls CreateFile. Testing with a different windows dd program gives more expected results. -- James Peach | [EMAIL PROTECTED] -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Fri, Apr 28, 2006 at 09:40:39AM +1000, James Peach wrote: On 4/27/06, Jeremy Allison [EMAIL PROTECTED] wrote: On Wed, Apr 26, 2006 at 01:20:23PM -0400, Greg Dickie wrote: OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? Greg - I LOVE YOU :-). That's almost certainly the problem. Back in the Win9x protocols days (the openX style of calls) we had the problem that Windows clients were setting the sync this data flag for almost every open - completely unneccessary. For 3.0.20x we changed to the NTcreateX style of open, and mapped the create option FILE_WRITE_THROUGH to the O_SYNC. I bet the Windows redirector in W2K, XP and W2K3 is just as dim, and is adding that on almost every open (I'm on a plane right now so can't get to my vmware sessions easily to check). xdd explicitly sets this flag when it calls CreateFile. Testing with a different windows dd program gives more expected results. The question is - what does netbench do ? And what does iometer do ? I'm back home now (but *very* tired) so I'll get some answers soon. It's definately something we needed to fix though. Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? Thanks, Greg Jeremy Allison wrote: On Mon, Apr 24, 2006 at 10:56:29AM -0400, Greg Dickie wrote: Hey Jeremy, Sorry, I'm at NAB in vegas and a little burned myself so we graph output from /proc/diskstats when we test. If I run a test with 3.0.21c the wio (field 7) and wblk (field 9) stats both show activity. This does not happen with 3.0.14a. eg: it appears that there is double the amount of write traffic. Does that make any more sense? Not really :-). I don't know enough about xfs to be dangerous :-). Is it possible we're doing larger bulk writes with 3.0.2x that we weren't doing with 3.0.14a which might trigger this ? In which case it'd be an xfs issue not a Samba one. Can you test with ext3 jfs or reiser to see if they show different performance characteristics ? Jeremy -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On 4/27/06, Greg Dickie [EMAIL PROTECTED] wrote: OK all my gobbledy-gook stats aside. sorry for being incoherent. Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could that be the source of the problems? I'll try and find that and test it when I get home but the question is is it necessary? O_SYNC is not a good option to use if you want good write performance on XFS. It's what you use to when you really want your data safe and don't care about throughput. -- James Peach | [EMAIL PROTECTED] -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Sun, Apr 23, 2006 at 10:24:37PM -0400, Greg Dickie wrote: Perhaps another data point that might ring a bell. While using 3.0.21c and monitoring direct IO vs. buffered IO (on XFS): if you see 40MBs of direct IO you will see another 40 MBs of buffered IO on 21c that you do not see on 14a. According to a colleague he saw this at one point in a previous revision and thought it was a problem of an extraneous sync somewhere This might explain why you don't see this with your ram disk. Ok, I don't understand this message :-). Can you explain exactly what you mean with 40 MBs of bufferd IO on 21c that you do not see on 14a - how are you measuring this and where do you see it ? Use small words please, I'm very jet-lagged :-). Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Hey Jeremy, Sorry, I'm at NAB in vegas and a little burned myself so we graph output from /proc/diskstats when we test. If I run a test with 3.0.21c the wio (field 7) and wblk (field 9) stats both show activity. This does not happen with 3.0.14a. eg: it appears that there is double the amount of write traffic. Does that make any more sense? Thanks, Greg Jeremy Allison wrote: On Sun, Apr 23, 2006 at 10:24:37PM -0400, Greg Dickie wrote: Perhaps another data point that might ring a bell. While using 3.0.21c and monitoring direct IO vs. buffered IO (on XFS): if you see 40MBs of direct IO you will see another 40 MBs of buffered IO on 21c that you do not see on 14a. According to a colleague he saw this at one point in a previous revision and thought it was a problem of an extraneous sync somewhere This might explain why you don't see this with your ram disk. Ok, I don't understand this message :-). Can you explain exactly what you mean with 40 MBs of bufferd IO on 21c that you do not see on 14a - how are you measuring this and where do you see it ? Use small words please, I'm very jet-lagged :-). Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Mon, Apr 24, 2006 at 10:56:29AM -0400, Greg Dickie wrote: Hey Jeremy, Sorry, I'm at NAB in vegas and a little burned myself so we graph output from /proc/diskstats when we test. If I run a test with 3.0.21c the wio (field 7) and wblk (field 9) stats both show activity. This does not happen with 3.0.14a. eg: it appears that there is double the amount of write traffic. Does that make any more sense? Not really :-). I don't know enough about xfs to be dangerous :-). Is it possible we're doing larger bulk writes with 3.0.2x that we weren't doing with 3.0.14a which might trigger this ? In which case it'd be an xfs issue not a Samba one. Can you test with ext3 jfs or reiser to see if they show different performance characteristics ? Jeremy -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Perhaps another data point that might ring a bell. While using 3.0.21c and monitoring direct IO vs. buffered IO (on XFS): if you see 40MBs of direct IO you will see another 40 MBs of buffered IO on 21c that you do not see on 14a. According to a colleague he saw this at one point in a previous revision and thought it was a problem of an extraneous sync somewhere This might explain why you don't see this with your ram disk. Thanks, Greg Greg Dickie wrote: So this is testing 14a and SVN on the exact same machine with the exact same configuration. The only difference is switching samba RPM. I wanted to get 14a numbers to cre-confirm the setup but unfortunately the KVM seems to have gone on strike. More news as soon as possible. Thanks, Greg On Wed, 2006-04-19 at 17:32 -0700, Jeremy Allison wrote: On Wed, Apr 19, 2006 at 07:26:16PM -0400, Greg Dickie wrote: Hi Jeremy, Bad news I'm afraid. Doesn't seem to be much of a difference between that svn checkout and 3.0.20. Thats just the first run on it, we'll try and poke it some more. Thanks alot for your work on this, No problem. But I'm testing here on a Linux ram disk with ext2 as a target to remove any possible variance caused by disk activity and with iometer get equal performance (within noise values) between 3.0.14a and SVN SAMBA_3_0. Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Wed, Apr 19, 2006 at 07:26:16PM -0400, Greg Dickie wrote: Bad news I'm afraid. Doesn't seem to be much of a difference between that svn checkout and 3.0.20. Thats just the first run on it, we'll try and poke it some more. One thing that had smoothed it for me is to use use spnego = no Please be aware that this is nothing I would recommend for production, but I'd be interested if setting that parameter also equalizes 3.0.14 and the latest code for you. Volker pgph9VAM4dgis.pgp Description: PGP signature -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Thanks Volker, we'll try that, at least that may indicate where the bottleneck is. Greg On Thu, 2006-04-20 at 10:35 +0200, Volker Lendecke wrote: On Wed, Apr 19, 2006 at 07:26:16PM -0400, Greg Dickie wrote: Bad news I'm afraid. Doesn't seem to be much of a difference between that svn checkout and 3.0.20. Thats just the first run on it, we'll try and poke it some more. One thing that had smoothed it for me is to use use spnego = no Please be aware that this is nothing I would recommend for production, but I'd be interested if setting that parameter also equalizes 3.0.14 and the latest code for you. Volker -- Greg Dickie just a guy Maximum Throughput -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
So this is testing 14a and SVN on the exact same machine with the exact same configuration. The only difference is switching samba RPM. I wanted to get 14a numbers to cre-confirm the setup but unfortunately the KVM seems to have gone on strike. More news as soon as possible. Thanks, Greg On Wed, 2006-04-19 at 17:32 -0700, Jeremy Allison wrote: On Wed, Apr 19, 2006 at 07:26:16PM -0400, Greg Dickie wrote: Hi Jeremy, Bad news I'm afraid. Doesn't seem to be much of a difference between that svn checkout and 3.0.20. Thats just the first run on it, we'll try and poke it some more. Thanks alot for your work on this, No problem. But I'm testing here on a Linux ram disk with ext2 as a target to remove any possible variance caused by disk activity and with iometer get equal performance (within noise values) between 3.0.14a and SVN SAMBA_3_0. Jeremy. -- Greg Dickie just a guy Maximum Throughput -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Hi Jeremy, Bad news I'm afraid. Doesn't seem to be much of a difference between that svn checkout and 3.0.20. Thats just the first run on it, we'll try and poke it some more. Thanks alot for your work on this, Greg On Tue, 2006-04-18 at 11:20 -0700, Jeremy Allison wrote: On Fri, Apr 07, 2006 at 01:05:43PM -0400, Greg Dickie wrote: Hey Jeremy, I think that just means 3 runs 1 reads, 1 writes and one 50-50. I could have sworn I sent the config file already but here it is again along with the binaries. Thanks alot, Greg Greg - is it possible for you to test the current code in the svn branch SAMBA_3_0 with this application and your test case ? I've been doing a lot of work on this issue (it's very important obviously :-) and would like to get some feedback if possible. Thanks, Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Wed, Apr 19, 2006 at 07:26:16PM -0400, Greg Dickie wrote: Hi Jeremy, Bad news I'm afraid. Doesn't seem to be much of a difference between that svn checkout and 3.0.20. Thats just the first run on it, we'll try and poke it some more. Thanks alot for your work on this, No problem. But I'm testing here on a Linux ram disk with ext2 as a target to remove any possible variance caused by disk activity and with iometer get equal performance (within noise values) between 3.0.14a and SVN SAMBA_3_0. Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Fri, Apr 07, 2006 at 01:05:43PM -0400, Greg Dickie wrote: Hey Jeremy, I think that just means 3 runs 1 reads, 1 writes and one 50-50. I could have sworn I sent the config file already but here it is again along with the binaries. Thanks alot, Greg Greg - is it possible for you to test the current code in the svn branch SAMBA_3_0 with this application and your test case ? I've been doing a lot of work on this issue (it's very important obviously :-) and would like to get some feedback if possible. Thanks, Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Absolutely! Checking it out and building now. news ASAP. Thanks alot, Greg On Tue, 2006-04-18 at 11:20 -0700, Jeremy Allison wrote: On Fri, Apr 07, 2006 at 01:05:43PM -0400, Greg Dickie wrote: Hey Jeremy, I think that just means 3 runs 1 reads, 1 writes and one 50-50. I could have sworn I sent the config file already but here it is again along with the binaries. Thanks alot, Greg Greg - is it possible for you to test the current code in the svn branch SAMBA_3_0 with this application and your test case ? I've been doing a lot of work on this issue (it's very important obviously :-) and would like to get some feedback if possible. Thanks, Jeremy. -- Greg Dickie just a guy Maximum Throughput -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Hi Jeremy, Just want to confirm that you received the iometer binary and config. I tried sending to the list but it never got posted so I tried just sending to you but maybe there is a hyper virus filter somewhere that blocked it... Greg Jeremy Allison wrote: On Fri, Mar 31, 2006 at 09:49:02AM -0500, Greg Dickie wrote: Yes, it is the win32 version, but its an old one (1998), not sure how the config file will carry to a more recent version. IOmeter Access specifications: Transfer req. size = 64Kb 8Kb Percent of Access Specification = 100% Type of Operation: 100% read; 100% write; 50% read Percent Random/Sequential Distribution: 100% sequential Queue Depth = 8 Reply Size: No Reply Burst Length = 25 I/Os Align I/Os on: 64Kb 8Kb Ramp up Time = 30sec Run Time = 3 min # of clients used: 1 to 9 clients (Linear Stepping) Ok, I'm trying to work out how to specify this but it doesn't make sense (this line in particular : Type of Operation: 100% read; 100% write; 50% read with the new version. Can you just attach the config file as an attachment and I'll see if it can read it ? (you could just send the 1998 binary for complete reproducibility). Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Sun, Apr 09, 2006 at 09:53:57AM -0400, Greg Dickie wrote: Hi Jeremy, Just want to confirm that you received the iometer binary and config. I tried sending to the list but it never got posted so I tried just sending to you but maybe there is a hyper virus filter somewhere that blocked it... Yes I got it. I'm hoping to work on it this week... Thanks, Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Hey Jeremy, I think that just means 3 runs 1 reads, 1 writes and one 50-50. I could have sworn I sent the config file already but here it is again along with the binaries. Thanks alot, Greg On Thu, 2006-04-06 at 18:33 -0700, Jeremy Allison wrote: On Fri, Mar 31, 2006 at 09:49:02AM -0500, Greg Dickie wrote: Yes, it is the win32 version, but its an old one (1998), not sure how the config file will carry to a more recent version. IOmeter Access specifications: Transfer req. size = 64Kb 8Kb Percent of Access Specification = 100% Type of Operation: 100% read; 100% write; 50% read Percent Random/Sequential Distribution: 100% sequential Queue Depth = 8 Reply Size: No Reply Burst Length = 25 I/Os Align I/Os on: 64Kb 8Kb Ramp up Time = 30sec Run Time = 3 min # of clients used: 1 to 9 clients (Linear Stepping) Ok, I'm trying to work out how to specify this but it doesn't make sense (this line in particular : Type of Operation: 100% read; 100% write; 50% read with the new version. Can you just attach the config file as an attachment and I'll see if it can read it ? (you could just send the 1998 binary for complete reproducibility). Jeremy. -- Greg Dickie just a guy Maximum Throughput 'Version 1998.10.08 'Access specifications 'Access specification name,default assignment Read 100% 64k,1 'size,% of size,% reads,% random,delay,burst,align,reply 65536,100,100,0,0,25,65536,0 'Access specification name,default assignment Write 100% 64k,1 'size,% of size,% reads,% random,delay,burst,align,reply 65536,100,0,0,0,25,65536,0 'Access specification name,default assignment Read 50% 64k,1 'size,% of size,% reads,% random,delay,burst,align,reply 65536,100,50,0,0,25,65536,0 'Access specification name,default assignment Read 100% 8k,1 'size,% of size,% reads,% random,delay,burst,align,reply 8192,100,100,0,0,25,8192,0 'Access specification name,default assignment Write 100% 8k,1 'size,% of size,% reads,% random,delay,burst,align,reply 8192,100,0,0,0,25,8192,0 'Access specification name,default assignment Read 50% 8k,1 'size,% of size,% reads,% random,delay,burst,align,reply 8192,100,50,0,0,25,8192,0 'End access specifications 'Test Setup 'Test Description 'Run Time ' hoursminutesseconds 0 3 0 'Ramp Up Time (ms) 30 'Default Disk Workers to Spawn -1 'Default Network Workers to Spawn 0 'Record Results 0 'Worker Cycling ' start step step type 1 1 0 'Disk Cycling ' start step step type 1 1 0 'Queue Depth Cycling ' startend step step type 1 32 2 1 'Test Type 4 'Version 1998.10.08 -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Fri, Mar 31, 2006 at 09:49:02AM -0500, Greg Dickie wrote: Yes, it is the win32 version, but its an old one (1998), not sure how the config file will carry to a more recent version. IOmeter Access specifications: Transfer req. size = 64Kb 8Kb Percent of Access Specification = 100% Type of Operation: 100% read; 100% write; 50% read Percent Random/Sequential Distribution: 100% sequential Queue Depth = 8 Reply Size: No Reply Burst Length = 25 I/Os Align I/Os on: 64Kb 8Kb Ramp up Time = 30sec Run Time = 3 min # of clients used: 1 to 9 clients (Linear Stepping) Ok, I'm trying to work out how to specify this but it doesn't make sense (this line in particular : Type of Operation: 100% read; 100% write; 50% read with the new version. Can you just attach the config file as an attachment and I'll see if it can read it ? (you could just send the 1998 binary for complete reproducibility). Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Yes, it is the win32 version, but its an old one (1998), not sure how the config file will carry to a more recent version. IOmeter Access specifications: Transfer req. size = 64Kb 8Kb Percent of Access Specification = 100% Type of Operation: 100% read; 100% write; 50% read Percent Random/Sequential Distribution: 100% sequential Queue Depth = 8 Reply Size: No Reply Burst Length = 25 I/Os Align I/Os on: 64Kb 8Kb Ramp up Time = 30sec Run Time = 3 min # of clients used: 1 to 9 clients (Linear Stepping) Note that the problem is visible with only one client. The clients are GbE as is the server, not sure if you will see a difference on 100BT. Thank you very much for looking at this, please let me know if there is anything I can do to help. Regards, Greg On Thu, 2006-03-30 at 17:20 -0800, Jeremy Allison wrote: On Thu, Mar 30, 2006 at 08:04:21PM -0500, Greg Dickie wrote: Hey Jeremy, This happens when testing under iometer. I can give you the exact paramaters if you want. That's the Win32 version ? If so, yes please the exact parameters would be extremely useful as I can reproduce this here. Thanks, Jeremy. -- Greg Dickie just a guy Maximum Throughput 'Version 1998.10.08 'Access specifications 'Access specification name,default assignment Read 100% 64k,1 'size,% of size,% reads,% random,delay,burst,align,reply 65536,100,100,0,0,25,65536,0 'Access specification name,default assignment Write 100% 64k,1 'size,% of size,% reads,% random,delay,burst,align,reply 65536,100,0,0,0,25,65536,0 'Access specification name,default assignment Read 50% 64k,1 'size,% of size,% reads,% random,delay,burst,align,reply 65536,100,50,0,0,25,65536,0 'Access specification name,default assignment Read 100% 8k,1 'size,% of size,% reads,% random,delay,burst,align,reply 8192,100,100,0,0,25,8192,0 'Access specification name,default assignment Write 100% 8k,1 'size,% of size,% reads,% random,delay,burst,align,reply 8192,100,0,0,0,25,8192,0 'Access specification name,default assignment Read 50% 8k,1 'size,% of size,% reads,% random,delay,burst,align,reply 8192,100,50,0,0,25,8192,0 'End access specifications 'Test Setup 'Test Description 'Run Time ' hoursminutesseconds 0 3 0 'Ramp Up Time (ms) 30 'Default Disk Workers to Spawn -1 'Default Network Workers to Spawn 0 'Record Results 0 'Worker Cycling ' start step step type 1 1 0 'Disk Cycling ' start step step type 1 1 0 'Queue Depth Cycling ' startend step step type 1 32 2 1 'Test Type 4 'Version 1998.10.08 -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Fri, Mar 31, 2006 at 09:49:02AM -0500, Greg Dickie wrote: Yes, it is the win32 version, but its an old one (1998), not sure how the config file will carry to a more recent version. IOmeter Access specifications: Transfer req. size = 64Kb 8Kb Percent of Access Specification = 100% Type of Operation: 100% read; 100% write; 50% read Percent Random/Sequential Distribution: 100% sequential Queue Depth = 8 Reply Size: No Reply Burst Length = 25 I/Os Align I/Os on: 64Kb 8Kb Ramp up Time = 30sec Run Time = 3 min # of clients used: 1 to 9 clients (Linear Stepping) Note that the problem is visible with only one client. The clients are GbE as is the server, not sure if you will see a difference on 100BT. Thank you very much for looking at this, please let me know if there is anything I can do to help. No problem. What I'll do is run the version against a smbd running with cachegrind. That will point out any extra CPU usage we're accumulating between the two versions. If we've regressed because of a code path this will tell us. Might take a while though as I have to be at LinuxWorld Boston next week. Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
I am unfamiliar with cachegrind but I assume this is just a profiler. Do you have to compile instrumentation into smbd? Have fun in Boston, great city. Greg On Fri, 2006-03-31 at 09:17 -0800, Jeremy Allison wrote: On Fri, Mar 31, 2006 at 09:49:02AM -0500, Greg Dickie wrote: Yes, it is the win32 version, but its an old one (1998), not sure how the config file will carry to a more recent version. IOmeter Access specifications: Transfer req. size = 64Kb 8Kb Percent of Access Specification = 100% Type of Operation: 100% read; 100% write; 50% read Percent Random/Sequential Distribution: 100% sequential Queue Depth = 8 Reply Size: No Reply Burst Length = 25 I/Os Align I/Os on: 64Kb 8Kb Ramp up Time = 30sec Run Time = 3 min # of clients used: 1 to 9 clients (Linear Stepping) Note that the problem is visible with only one client. The clients are GbE as is the server, not sure if you will see a difference on 100BT. Thank you very much for looking at this, please let me know if there is anything I can do to help. No problem. What I'll do is run the version against a smbd running with cachegrind. That will point out any extra CPU usage we're accumulating between the two versions. If we've regressed because of a code path this will tell us. Might take a while though as I have to be at LinuxWorld Boston next week. Jeremy. -- Greg Dickie just a guy Maximum Throughput -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Fri, Mar 31, 2006 at 02:32:44PM -0500, Greg Dickie wrote: I am unfamiliar with cachegrind but I assume this is just a profiler. Do you have to compile instrumentation into smbd? No, that's why it's so powerful. It's part of the valgrind suite, which emulates a complete x86 cpu as an LD_PRELOAD. You get *complete* information about all CPU activity, even though it runs 10x slower than usual (but I only care about relative changes so that's not an issue here). Have fun in Boston, great city. Thanks, Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
Hey Jeremy, This happens when testing under iometer. I can give you the exact paramaters if you want. Thanks, Greg On Wed, 2006-03-29 at 09:50 -0800, Jeremy Allison wrote: On Wed, Mar 29, 2006 at 11:03:02AM -0500, Greg Dickie wrote: Suddenly occurred to me that I wasn't seeing the performance I used to while running 3.0.21c. Went back and tested 3.0.14a on the exact same configuration and boom, smaller writes (8K) were about twice as fast. I narrowed it down to a change between 14a and 20 but there were alot of changes in there. Can anyone think of what this could be? How are you testing this ? If you can give me the exact same test config I can run under cachegrind and it'll tell me exactly where the extra time is being spent between the 3.0.14a and 3.0.20 versions. Thanks, Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Thu, Mar 30, 2006 at 08:04:21PM -0500, Greg Dickie wrote: Hey Jeremy, This happens when testing under iometer. I can give you the exact paramaters if you want. That's the Win32 version ? If so, yes please the exact parameters would be extremely useful as I can reproduce this here. Thanks, Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba
Re: [Samba] performance regression between 30.14a and 3.0.20
On Wed, Mar 29, 2006 at 11:03:02AM -0500, Greg Dickie wrote: Suddenly occurred to me that I wasn't seeing the performance I used to while running 3.0.21c. Went back and tested 3.0.14a on the exact same configuration and boom, smaller writes (8K) were about twice as fast. I narrowed it down to a change between 14a and 20 but there were alot of changes in there. Can anyone think of what this could be? How are you testing this ? If you can give me the exact same test config I can run under cachegrind and it'll tell me exactly where the extra time is being spent between the 3.0.14a and 3.0.20 versions. Thanks, Jeremy. -- To unsubscribe from this list go to the following URL and read the instructions: https://lists.samba.org/mailman/listinfo/samba