Re: Performance of -current vs -stable
Serwoas! %s wrote on %.3s, %lld Sep 1993 > > Could it be due to the DDB, INVARIANTS & WITNESS options in the > > kernel? If it is that's fine with me, I'm just wondering where > > that magnitude of a slowdown would be coming from. > WITNESS can really hurt. Quite possibly I should turn it off in GENERIC now (I Hmmm, a few weeks ago I did some totally unscientific testing, noting that -current was much slower than -stable, by playing an mp3 with an optimized `mpg123' on a 75MHz pentium machine, where the difference was far more obvious than on a faster machine. I suspected that WITNESS and similar goodies might be a problem, and even composed a lengthy message wondering about it, but what really got me wondering was the fact that -stable had far less idle time than the same hardware running NetBSD-current, so that I could do `useful' work under NetBSD while playing an mp3 cleanly, but such was difficult with FreeBSD-stable and impossible with -current. However, I suspect that such a question is more on-topic in -hackers or even -stable than here, but I'm wondering if I should extract any useful info from the message I composed but never sent, and post my kernel config and ask if there's anything obvious in there that would explain why FreeBSD's mpg123 takes ~60% CPU and NetBSD's ~30% (vs the ~90+% usage by -current)... Oh, I'll try rebuilding -current Real Soon^W^W later today, without WITNESS, and compare, just to stay on-topic for this list. thanks barry bouwsma To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On 07-Feb-02 Georg-W Koltermann wrote: > At Wed, 06 Feb 2002 23:33:14 -0500 (EST), > John Baldwin wrote: >> >> [...] >> I guess. Note that you can use a loader tunable 'debug.witness_watch' to >> turn >> witness off from the loader. If it's set to 0 witness won't be used even if >> it's compiled into the kernel (just a general FYI, witness(4) documents this >> as >> well). > > Nice feature, thanks for the hint. I can't find it documented in > witness(4), however: > > hunter# grep -i watch /usr/src/share/man/man4/witness.4 > hunter# > > CVSup is from Feb 6. Hmm, doh. I document the other loader tunables and sysctls, just not debug.witness_watch. If someone with some mdoc fu and spare time wants to add a para noting that debug.witness_watch can be set to 0 as a loader tunable to turn off witness and that the read-only sysctl debug.witness_watch can be used at runtime to see if witness is enabled or not. > -- > Regards, > Georg. -- John Baldwin <[EMAIL PROTECTED]> <>< http://www.FreeBSD.org/~jhb/ "Power Users Use the Power to Serve!" - http://www.FreeBSD.org/ To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Wed, Feb 06, 2002 at 12:10:05AM -0500, Garance A Drosihn wrote: > On current On stable > -- -- > real7m 43.392s 4m 53.100sin /usr/src for current > user0m 11.692s 0m 4.203s > sys 3m 4.601s 0m 2.248s > > real6m 40.322s 2m 39.361sin /usr/src for stable > user0m 10.531s 0m 6.653s > sys 4m 28.863s 0m 9.480s ... > Could it be due to the DDB, INVARIANTS & WITNESS options in the The order of magnitude increase in 'sys' time is the indication that WITNESS is turned on. Infact IIRC I once did a 'make -j8 buildworld' on a dual Athlon system, and had sys > user. -- -- David ([EMAIL PROTECTED]) To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
At Wed, 06 Feb 2002 23:33:14 -0500 (EST), John Baldwin wrote: > > [...] > I guess. Note that you can use a loader tunable 'debug.witness_watch' to turn > witness off from the loader. If it's set to 0 witness won't be used even if > it's compiled into the kernel (just a general FYI, witness(4) documents this as > well). Nice feature, thanks for the hint. I can't find it documented in witness(4), however: hunter# grep -i watch /usr/src/share/man/man4/witness.4 hunter# CVSup is from Feb 6. -- Regards, Georg. To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On 6 Feb, Garance A Drosihn wrote: > Anything else I should check? I realize there's about a million > differences between the two branches, and there might also be > something about my machine's setup which is a major culprit here. > I'm just looking for a basic idea of what other people have been > seeing for performance when they run current. # ll /etc/malloc.conf lrwx-- 1 root wheel 2B Aug 18 21:47 /etc/malloc.conf@ -> aj Bye, Alexander. -- Loose bits sink chips. http://www.Leidinger.net Alexander @ Leidinger.net GPG fingerprint = C518 BC70 E67F 143F BE91 3365 79E2 9C60 B006 3FE7 To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Thu, 7 Feb 2002 19:08:07 +0100 Wilko Bulte <[EMAIL PROTECTED]> wrote: > On Thu, Feb 07, 2002 at 10:15:02AM +0100, Cejka Rudolf wrote: > > > I'm just looking for a basic idea of what other people have been > > > seeing for performance when they run current. > > > > There is another common source of confusion: If anybody has IDE > > disks, write-caching is enabled by default in -stable, but disabled > > in -current. > > I don't think that is true anymore. -stable has WC enabled as well. -STABLE has it enabled indeed, but -CURRENT does not (afaik). Read what he said. :-) > > -- > | / o / /_ _ [EMAIL PROTECTED] > |/|/ / / /( (_) Bulte Arnhem, the Netherlands > > To Unsubscribe: send mail to [EMAIL PROTECTED] > with "unsubscribe freebsd-current" in the body of the message > To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Thu, Feb 07, 2002 at 10:15:02AM +0100, Cejka Rudolf wrote: > Garance A Drosihn wrote (2002/02/06): > > Anything else I should check? I realize there's about a million > > differences between the two branches, and there might also be > > something about my machine's setup which is a major culprit here. > > I'm just looking for a basic idea of what other people have been > > seeing for performance when they run current. > > There is another common source of confusion: If anybody has IDE > disks, write-caching is enabled by default in -stable, but disabled > in -current. I don't think that is true anymore. -stable has WC enabled as well. -- | / o / /_ _ [EMAIL PROTECTED] |/|/ / / /( (_) Bulte Arnhem, the Netherlands To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Performance of -current vs -stable
With 4.5-release out the door, I thought I'd start trying to use 5.0-current on my "main freebsd machine" instead of 4.x-stable. I figure at some point we (as developers) have got to try to migrate to that release as much as possible. I had been doing some stuff with 5.0-current at home. That seemed a bit slow, but I didn't think too much of it as the home machine is a single-CPU Duron-based machine (600 Mhz). while the machine at work is a dual-CPU Pentium-3 machine (650 MHz). The office machine also has more RAM, so obviously the home machine would be slower. But switching to current on the office machine is also considerably slower. I wasn't expecting "faster", but I'm wondering if other people are also seeing it as much slower, or if I've just got some other odd problem hitting me. One simple test I tried was that I have a copy of the freebsd cvs repository in /usr/cvs/free, on it's own partition. Each system has it's own /usr/src, of course. I cvsup'ed /usr/cvs/free, and then did a time cvs status >/dev/null in each /usr/src, on each system. (that command is just for this timing test, I usually do more useful commands at that point!). On current On stable -- -- real7m 43.392s 4m 53.100sin /usr/src for current user0m 11.692s 0m 4.203s sys 3m 4.601s 0m 2.248s real6m 40.322s 2m 39.361sin /usr/src for stable user0m 10.531s 0m 6.653s sys 4m 28.863s 0m 9.480s I realize this example isn't terribly detailed, but it seems to match what I "feel" when doing any major work on current. I'm used to a 'make -j5 buildworld' taking between 42 and 45 minutes for stable on my office machine, but the few times I've rebuilt -current, it's taking more like three and a half hours. That is quite a hit. This current system was initially installed in late-december, doing a full system-install (booting off a CD, newfs'ing the partitions, etc). So, it should have picked up all the recent filesystem improvements (dirprefs, etc) when laying out the files, and even if that wasn't true this test is done using the exact same directories as source & destination for the stable vs current tests. Could it be due to the DDB, INVARIANTS & WITNESS options in the kernel? If it is that's fine with me, I'm just wondering where that magnitude of a slowdown would be coming from. Anything else I should check? I realize there's about a million differences between the two branches, and there might also be something about my machine's setup which is a major culprit here. I'm just looking for a basic idea of what other people have been seeing for performance when they run current. -- Garance Alistair Drosehn= [EMAIL PROTECTED] Senior Systems Programmer or [EMAIL PROTECTED] Rensselaer Polytechnic Instituteor [EMAIL PROTECTED] To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
At 12:10 AM -0500 2/6/02, Garance A Drosihn wrote: >One simple test I tried was that I have a copy of the freebsd cvs >repository in /usr/cvs/free, on it's own partition. Each system >has it's own /usr/src, of course. I cvsup'ed /usr/cvs/free, and >then did a > time cvs status >/dev/null >in each /usr/src, on each system. The popular vote fingered the WITNESS option as the culprit for the significant slowdown I noticed in -current. I compiled a kernel which did nothing but turn that off, and here is an updated chart of my simple-minded benchmark: On current On currentOn stable +witness NO witness (no witness) -- -- -- real7m 43.392s 5m 9.443s 4m 53.100sin the /usr/src user0m 11.692s 0m 11.223s 0m 4.203s for current sys 3m 4.601s 0m 17.353s 0m 2.248s real6m 40.322s 1m 57.438s 2m 39.361sin the /usr/src user0m 10.531s 0m 9.474s 0m 6.653s for stable sys 4m 28.863s 0m 12.527s 0m 9.480s So, things are quite reasonable for my purposes if I turn off the WITNESS kernel option, and leave all the other debugging stuff active. reminder: The above is obviously not a scientific benchmark! It was just something adequate enough to show a rough idea of the performance I was "feeling". Thanks for the pointers. Now I'll try recompiling the rest of the ports I have on my -stable system, and see if I can't keep my office machine running in -current instead of -stable... -- Garance Alistair Drosehn= [EMAIL PROTECTED] Senior Systems Programmer or [EMAIL PROTECTED] Rensselaer Polytechnic Instituteor [EMAIL PROTECTED] To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On 07-Feb-02 Georg-W Koltermann wrote: > At Wed, 06 Feb 2002 23:33:14 -0500 (EST), > John Baldwin wrote: >> >> [...] >> I guess. Note that you can use a loader tunable 'debug.witness_watch' to >> turn >> witness off from the loader. If it's set to 0 witness won't be used even if >> it's compiled into the kernel (just a general FYI, witness(4) documents this >> as >> well). > > Nice feature, thanks for the hint. I can't find it documented in > witness(4), however: > > hunter# grep -i watch /usr/src/share/man/man4/witness.4 > hunter# > > CVSup is from Feb 6. Hmm, doh. I document the other loader tunables and sysctls, just not debug.witness_watch. If someone with some mdoc fu and spare time wants to add a para noting that debug.witness_watch can be set to 0 as a loader tunable to turn off witness and that the read-only sysctl debug.witness_watch can be used at runtime to see if witness is enabled or not. > -- > Regards, > Georg. -- John Baldwin <[EMAIL PROTECTED]> <>< http://www.FreeBSD.org/~jhb/ "Power Users Use the Power to Serve!" - http://www.FreeBSD.org/ To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
Wilko Bulte wrote: > > There is another common source of confusion: If anybody has IDE > > disks, write-caching is enabled by default in -stable, but disabled > > in -current. > > I don't think that is true anymore. -stable has WC enabled as well. "Friends don't let friends cache writes" ??? [1/2] 8-)... some IDE drives lie, when you tell them to flush their caches. [1] 8-(. -- Terry To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
> On Thu, 7 Feb 2002 19:08:07 +0100 > Wilko Bulte <[EMAIL PROTECTED]> wrote: > > > There is another common source of confusion: If anybody has IDE > > > disks, write-caching is enabled by default in -stable, but disabled > > > in -current. > > > > I don't think that is true anymore. -stable has WC enabled as well. > > -STABLE has it enabled indeed, but -CURRENT does not (afaik). > Read what he said. :-) I believe that hw.ata.wc is set to '0' by default. Easily correctible either by adding 'hw.ata.wc="1"' into your /boot/loader.conf or tweaking the source in /usr/src/sys/dev/ata/ata-disk.c (the former being the preferred method of course). -Steve To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Thu, 7 Feb 2002 19:08:07 +0100 Wilko Bulte <[EMAIL PROTECTED]> wrote: > On Thu, Feb 07, 2002 at 10:15:02AM +0100, Cejka Rudolf wrote: > > > I'm just looking for a basic idea of what other people have been > > > seeing for performance when they run current. > > > > There is another common source of confusion: If anybody has IDE > > disks, write-caching is enabled by default in -stable, but disabled > > in -current. > > I don't think that is true anymore. -stable has WC enabled as well. -STABLE has it enabled indeed, but -CURRENT does not (afaik). Read what he said. :-) > > -- > | / o / /_ _ [EMAIL PROTECTED] > |/|/ / / /( (_) Bulte Arnhem, the Netherlands > > To Unsubscribe: send mail to [EMAIL PROTECTED] > with "unsubscribe freebsd-current" in the body of the message > To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Thu, Feb 07, 2002 at 10:15:02AM +0100, Cejka Rudolf wrote: > Garance A Drosihn wrote (2002/02/06): > > Anything else I should check? I realize there's about a million > > differences between the two branches, and there might also be > > something about my machine's setup which is a major culprit here. > > I'm just looking for a basic idea of what other people have been > > seeing for performance when they run current. > > There is another common source of confusion: If anybody has IDE > disks, write-caching is enabled by default in -stable, but disabled > in -current. I don't think that is true anymore. -stable has WC enabled as well. -- | / o / /_ _ [EMAIL PROTECTED] |/|/ / / /( (_) Bulte Arnhem, the Netherlands To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
At Wed, 06 Feb 2002 23:33:14 -0500 (EST), John Baldwin wrote: > > [...] > I guess. Note that you can use a loader tunable 'debug.witness_watch' to turn > witness off from the loader. If it's set to 0 witness won't be used even if > it's compiled into the kernel (just a general FYI, witness(4) documents this as > well). Nice feature, thanks for the hint. I can't find it documented in witness(4), however: hunter# grep -i watch /usr/src/share/man/man4/witness.4 hunter# CVSup is from Feb 6. -- Regards, Georg. To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
Garance A Drosihn wrote (2002/02/06): > Anything else I should check? I realize there's about a million > differences between the two branches, and there might also be > something about my machine's setup which is a major culprit here. > I'm just looking for a basic idea of what other people have been > seeing for performance when they run current. There is another common source of confusion: If anybody has IDE disks, write-caching is enabled by default in -stable, but disabled in -current. -- Rudolf Cejka http://www.fit.vutbr.cz/~cejkar Brno University of Technology, Faculty of Information Technology Bozetechova 2, 612 66 Brno, Czech Republic To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On 06-Feb-02 David O'Brien wrote: > On Wed, Feb 06, 2002 at 01:02:34AM -0500, John Baldwin wrote: >> WITNESS can really hurt. Quite possibly I should turn it off in >> GENERIC now (I wouldn't mind if someone else did that.) > > I think it should stay. Especially as we are not getting much usage in > -CURRENT. If we turn it off by default, it should come back on 3 mo. > before 5.0-RELEASE for testing. (and yes off for the actual release). I guess. Note that you can use a loader tunable 'debug.witness_watch' to turn witness off from the loader. If it's set to 0 witness won't be used even if it's compiled into the kernel (just a general FYI, witness(4) documents this as well). -- John Baldwin <[EMAIL PROTECTED]> <>< http://www.FreeBSD.org/~jhb/ "Power Users Use the Power to Serve!" - http://www.FreeBSD.org/ To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
At 12:10 AM -0500 2/6/02, Garance A Drosihn wrote: >One simple test I tried was that I have a copy of the freebsd cvs >repository in /usr/cvs/free, on it's own partition. Each system >has it's own /usr/src, of course. I cvsup'ed /usr/cvs/free, and >then did a > time cvs status >/dev/null >in each /usr/src, on each system. The popular vote fingered the WITNESS option as the culprit for the significant slowdown I noticed in -current. I compiled a kernel which did nothing but turn that off, and here is an updated chart of my simple-minded benchmark: On current On currentOn stable +witness NO witness (no witness) -- -- -- real7m 43.392s 5m 9.443s 4m 53.100sin the /usr/src user0m 11.692s 0m 11.223s 0m 4.203s for current sys 3m 4.601s 0m 17.353s 0m 2.248s real6m 40.322s 1m 57.438s 2m 39.361sin the /usr/src user0m 10.531s 0m 9.474s 0m 6.653s for stable sys 4m 28.863s 0m 12.527s 0m 9.480s So, things are quite reasonable for my purposes if I turn off the WITNESS kernel option, and leave all the other debugging stuff active. reminder: The above is obviously not a scientific benchmark! It was just something adequate enough to show a rough idea of the performance I was "feeling". Thanks for the pointers. Now I'll try recompiling the rest of the ports I have on my -stable system, and see if I can't keep my office machine running in -current instead of -stable... -- Garance Alistair Drosehn= [EMAIL PROTECTED] Senior Systems Programmer or [EMAIL PROTECTED] Rensselaer Polytechnic Instituteor [EMAIL PROTECTED] To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
At 9:13 AM -0800 2/6/02, David O'Brien wrote: >On Wed, Feb 06, 2002 at 01:02:34AM -0500, John Baldwin wrote: >> WITNESS can really hurt. Quite possibly I should turn it off in >> GENERIC now (I wouldn't mind if someone else did that.) > >I think it should stay. Especially as we are not getting much usage in >-CURRENT. If we turn it off by default, it should come back on 3 mo. >before 5.0-RELEASE for testing. (and yes off for the actual release). I was thinking that maybe it should stay in the GENERIC kernel, along with some comment that indicates how much of a performance hit will be seen if it's on. "This is desirable for in-depth debugging, but it does mean that system calls will take three times longer than if the option is off". It's good to have the extra checking, but we don't want to scare people away from running current simply because they think it will take three times longer to get anything done on current. (or maybe even comment the option out, but do leave the line there along with the statement that "we'd appreciate it if people would run with this option on, even though it will cause a noticeable slowdown."). I do think it's important that we (developers) spend more time on current, but I think the way I'll do it is to compile two kernels, one with the WITNESS and one without it. For much of what I do I can absorb the extra overhead of WITNESS code, but a three-hour buildworld will throw my entire "buildworld schedule" off, and I just won't be able to rebuild the system as often. -- Garance Alistair Drosehn= [EMAIL PROTECTED] Senior Systems Programmer or [EMAIL PROTECTED] Rensselaer Polytechnic Instituteor [EMAIL PROTECTED] To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Wednesday 06 February 2002 06:10, Garance A Drosihn wrote: [performance complaint snip] > Could it be due to the DDB, INVARIANTS & WITNESS options in the > kernel? If it is that's fine with me, I'm just wondering where > that magnitude of a slowdown would be coming from. I would think so (iow, those are the biggest culprits). There are also some other issues (wrt locking and probably some other stuff) but they are bearable. I am running a PII 400 with 64 MB with CURRENT, it mighn't be the fastest now, but at least it works. I have been stresstesting the damn thing (as in, run lots of (relatively) useless services on there like icecast, and more ;-) and it is still holding. So far so good. The performance is low? Didn't notice that (but do keep in mind I'm happy with very little :-). It's not even buckling under a constant load avg of between 1.something and 3.something and 200+ procs running at all times... No dataloss yet. Very good. [snip] > I'm just looking for a basic idea of what other people have been > seeing for performance when they run current. My performance is okay from what I expect of it. Of course I have DDB, INVARIANTS and WITNESS options on (to provide better info for the developers when it crashes for some unexplainable reason). If you are planning to do 'production' (mind the quotes) stuff with CURRENT, I suggest you could strip those. But beware, you are on the bleeding edge here :-) Cheers, Emiel -- The problem with the gene pool is that there is no lifeguard. To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Wed, Feb 06, 2002 at 12:10:05AM -0500, Garance A Drosihn wrote: > On current On stable > -- -- > real7m 43.392s 4m 53.100sin /usr/src for current > user0m 11.692s 0m 4.203s > sys 3m 4.601s 0m 2.248s > > real6m 40.322s 2m 39.361sin /usr/src for stable > user0m 10.531s 0m 6.653s > sys 4m 28.863s 0m 9.480s ... > Could it be due to the DDB, INVARIANTS & WITNESS options in the The order of magnitude increase in 'sys' time is the indication that WITNESS is turned on. Infact IIRC I once did a 'make -j8 buildworld' on a dual Athlon system, and had sys > user. -- -- David ([EMAIL PROTECTED]) To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Wed, Feb 06, 2002 at 09:40:50AM -0500, Steve Ames wrote: > Is 'AJ' still set in the malloc options? I seem to recall setting > /etc/malloc.conf once in CURRENT for performance reasons. Yes it still is. That is the other thing Garance needs to decide if he wants on or off. To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Wed, Feb 06, 2002 at 01:02:34AM -0500, John Baldwin wrote: > WITNESS can really hurt. Quite possibly I should turn it off in > GENERIC now (I wouldn't mind if someone else did that.) I think it should stay. Especially as we are not getting much usage in -CURRENT. If we turn it off by default, it should come back on 3 mo. before 5.0-RELEASE for testing. (and yes off for the actual release). To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On Wed, Feb 06, 2002 at 01:02:34AM -0500, John Baldwin wrote: > > On 06-Feb-02 Garance A Drosihn wrote: > > Could it be due to the DDB, INVARIANTS & WITNESS options in the > > kernel? If it is that's fine with me, I'm just wondering where > > that magnitude of a slowdown would be coming from. > > WITNESS can really hurt. Quite possibly I should turn it off in GENERIC > now (I wouldn't mind if someone else did that.) Before major locking > changes people should still use it to look for bugs, but it isn't > very efficient I'm afraid. :( Is 'AJ' still set in the malloc options? I seem to recall setting /etc/malloc.conf once in CURRENT for performance reasons. -Steve To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
On 6 Feb, Garance A Drosihn wrote: > Anything else I should check? I realize there's about a million > differences between the two branches, and there might also be > something about my machine's setup which is a major culprit here. > I'm just looking for a basic idea of what other people have been > seeing for performance when they run current. # ll /etc/malloc.conf lrwx-- 1 root wheel 2B Aug 18 21:47 /etc/malloc.conf@ -> aj Bye, Alexander. -- Loose bits sink chips. http://www.Leidinger.net Alexander @ Leidinger.net GPG fingerprint = C518 BC70 E67F 143F BE91 3365 79E2 9C60 B006 3FE7 To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
Serwoas! %s wrote on %.3s, %lld Sep 1993 > > Could it be due to the DDB, INVARIANTS & WITNESS options in the > > kernel? If it is that's fine with me, I'm just wondering where > > that magnitude of a slowdown would be coming from. > WITNESS can really hurt. Quite possibly I should turn it off in GENERIC now (I Hmmm, a few weeks ago I did some totally unscientific testing, noting that -current was much slower than -stable, by playing an mp3 with an optimized `mpg123' on a 75MHz pentium machine, where the difference was far more obvious than on a faster machine. I suspected that WITNESS and similar goodies might be a problem, and even composed a lengthy message wondering about it, but what really got me wondering was the fact that -stable had far less idle time than the same hardware running NetBSD-current, so that I could do `useful' work under NetBSD while playing an mp3 cleanly, but such was difficult with FreeBSD-stable and impossible with -current. However, I suspect that such a question is more on-topic in -hackers or even -stable than here, but I'm wondering if I should extract any useful info from the message I composed but never sent, and post my kernel config and ask if there's anything obvious in there that would explain why FreeBSD's mpg123 takes ~60% CPU and NetBSD's ~30% (vs the ~90+% usage by -current)... Oh, I'll try rebuilding -current Real Soon^W^W later today, without WITNESS, and compare, just to stay on-topic for this list. thanks barry bouwsma To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
RE: Performance of -current vs -stable
On 06-Feb-02 Garance A Drosihn wrote: > Could it be due to the DDB, INVARIANTS & WITNESS options in the > kernel? If it is that's fine with me, I'm just wondering where > that magnitude of a slowdown would be coming from. WITNESS can really hurt. Quite possibly I should turn it off in GENERIC now (I wouldn't mind if someone else did that.) Before major locking changes people should still use it to look for bugs, but it isn't very efficient I'm afraid. :( -- John Baldwin <[EMAIL PROTECTED]> <>< http://www.FreeBSD.org/~jhb/ "Power Users Use the Power to Serve!" - http://www.FreeBSD.org/ To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
I asked BDE about the same thing off-list: Yes, the debug options (INVARIANTS and WITNESS, especially the latter) slow down the kernel by a factor of 10 or so. Only progams that don't make many syscalls run reasonably fast. I only turn on these options for debugging (not often). Without them, syscalls should be an average of only (sigh) about 10% slower in -current. --- "M. Warner Losh" <[EMAIL PROTECTED]> wrote: > In message:> Garance A Drosihn <[EMAIL PROTECTED]> > writes: > : Could it be due to the DDB, INVARIANTS & WITNESS > options in the > : kernel? If it is that's fine with me, I'm just > wondering where > : that magnitude of a slowdown would be coming from. > > No way for DDB, Maybe for Invariants, almost > certainly for witness. > WITNESS is slow and tends to exagerate lock > contention stalls (or so > it seems). > > Warner > > To Unsubscribe: send mail to [EMAIL PROTECTED] > with "unsubscribe freebsd-current" in the body of > the message > > __ Do You Yahoo!? Send FREE Valentine eCards with Yahoo! Greetings! http://greetings.yahoo.com To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Re: Performance of -current vs -stable
In message:Garance A Drosihn <[EMAIL PROTECTED]> writes: : Could it be due to the DDB, INVARIANTS & WITNESS options in the : kernel? If it is that's fine with me, I'm just wondering where : that magnitude of a slowdown would be coming from. No way for DDB, Maybe for Invariants, almost certainly for witness. WITNESS is slow and tends to exagerate lock contention stalls (or so it seems). Warner To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
Performance of -current vs -stable
With 4.5-release out the door, I thought I'd start trying to use 5.0-current on my "main freebsd machine" instead of 4.x-stable. I figure at some point we (as developers) have got to try to migrate to that release as much as possible. I had been doing some stuff with 5.0-current at home. That seemed a bit slow, but I didn't think too much of it as the home machine is a single-CPU Duron-based machine (600 Mhz). while the machine at work is a dual-CPU Pentium-3 machine (650 MHz). The office machine also has more RAM, so obviously the home machine would be slower. But switching to current on the office machine is also considerably slower. I wasn't expecting "faster", but I'm wondering if other people are also seeing it as much slower, or if I've just got some other odd problem hitting me. One simple test I tried was that I have a copy of the freebsd cvs repository in /usr/cvs/free, on it's own partition. Each system has it's own /usr/src, of course. I cvsup'ed /usr/cvs/free, and then did a time cvs status >/dev/null in each /usr/src, on each system. (that command is just for this timing test, I usually do more useful commands at that point!). On current On stable -- -- real7m 43.392s 4m 53.100sin /usr/src for current user0m 11.692s 0m 4.203s sys 3m 4.601s 0m 2.248s real6m 40.322s 2m 39.361sin /usr/src for stable user0m 10.531s 0m 6.653s sys 4m 28.863s 0m 9.480s I realize this example isn't terribly detailed, but it seems to match what I "feel" when doing any major work on current. I'm used to a 'make -j5 buildworld' taking between 42 and 45 minutes for stable on my office machine, but the few times I've rebuilt -current, it's taking more like three and a half hours. That is quite a hit. This current system was initially installed in late-december, doing a full system-install (booting off a CD, newfs'ing the partitions, etc). So, it should have picked up all the recent filesystem improvements (dirprefs, etc) when laying out the files, and even if that wasn't true this test is done using the exact same directories as source & destination for the stable vs current tests. Could it be due to the DDB, INVARIANTS & WITNESS options in the kernel? If it is that's fine with me, I'm just wondering where that magnitude of a slowdown would be coming from. Anything else I should check? I realize there's about a million differences between the two branches, and there might also be something about my machine's setup which is a major culprit here. I'm just looking for a basic idea of what other people have been seeing for performance when they run current. -- Garance Alistair Drosehn= [EMAIL PROTECTED] Senior Systems Programmer or [EMAIL PROTECTED] Rensselaer Polytechnic Instituteor [EMAIL PROTECTED] To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message