Bug#581907: sysv-rc: performance regression with CONCURRENCY=makefile
Hello, > I notice Xorg starts after 27 seconds with parallel booting, and 40 > seconds with sequencial boot. This make me suspect the regression is > really an artifact of how you measure the boot duration, and not a > result of the real boot speed slowdown. Reading thid thread, I would be enclined to believe that too. > Kel is working on a new bootchart package, which make it a bit easier > to measure it accurately. With this, we look for when the CPU settles > down after boot. Did that give anything? This investigation into a possible regression hasn't been conclusive and isn't likely to get anywhere further now, but I wanted to check with you. Shall we close this? -- Pierre Ynard
Bug#581907: sysv-rc: performance regression with CONCURRENCY=makefile
On Wed, May 19, 2010 at 04:17, Petter Reinholdtsen wrote: > [Felipe Sateler] >> Hmm. I modified S07bootchart to wait in a semi-busy loop until a >> process called gnome-panel appears. Probably in the -none case since >> gdm starts last, gnome-panel is picked right after it starts so it >> doesn't show up. > > I am not sure that is what happening. I suspect gdm finishes several > seconds after the last init.d script exits. Gdm finishes, but gnome has not finished loading yet. > >> Well, the problem is defining what is actually boot length. While >> the actual boot time may be longer, it certainly feels a bit faster >> because gdm starts earlier. > > Right. I have tried to define what should be measured at > http://wiki.debian.org/ReleaseGoals/BootPerformance>. We should > measure the time it take from grub starts to the desktop is ready for > use when autologin is in effect. Well, that's what I'm trying to measure but don't really know how. > > I still suspect the parallel boot is faster, but the non-parallel boot > fail to measure gdm startup as it is happening after the last init.d > script terminates. > >> Do you have an idea on how to accurately measure the boot time? > > Kel is working on a new bootchart package, which make it a bit easier > to measure it accurately. With this, we look for when the CPU settles > down after boot. That would be an improvement over what I'm measuring right now. -- Saludos, Felipe Sateler -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
Bug#581907: sysv-rc: performance regression with CONCURRENCY=makefile
[Felipe Sateler] > Hmm. I modified S07bootchart to wait in a semi-busy loop until a > process called gnome-panel appears. Probably in the -none case since > gdm starts last, gnome-panel is picked right after it starts so it > doesn't show up. I am not sure that is what happening. I suspect gdm finishes several seconds after the last init.d script exits. > Well, the problem is defining what is actually boot length. While > the actual boot time may be longer, it certainly feels a bit faster > because gdm starts earlier. Right. I have tried to define what should be measured at http://wiki.debian.org/ReleaseGoals/BootPerformance>. We should measure the time it take from grub starts to the desktop is ready for use when autologin is in effect. I still suspect the parallel boot is faster, but the non-parallel boot fail to measure gdm startup as it is happening after the last init.d script terminates. > Do you have an idea on how to accurately measure the boot time? Kel is working on a new bootchart package, which make it a bit easier to measure it accurately. With this, we look for when the CPU settles down after boot. Happy hacking, -- Petter Reinholdtsen -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
Bug#581907: sysv-rc: performance regression with CONCURRENCY=makefile
On Tue, May 18, 2010 at 14:56, Petter Reinholdtsen wrote: > [Felipe Sateler] >> I have now stopped bootchart when gnome-panel is loaded. It is not a >> perfect hint, but should be closer than the earlier estimates. >> Times now seem closer but makefile is still slower. > > Thank you. > > I fail to find gnome-panel in bootchart-none-autologin.png. When was > it started? Hmm. I modified S07bootchart to wait in a semi-busy loop until a process called gnome-panel appears. Probably in the -none case since gdm starts last, gnome-panel is picked right after it starts so it doesn't show up. > > I notice Xorg starts after 27 seconds with parallel booting, and 40 > seconds with sequencial boot. This make me suspect the regression is > really an artifact of how you measure the boot duration, and not a > result of the real boot speed slowdown. Well, the problem is defining what is actually boot length. While the actual boot time may be longer, it certainly feels a bit faster because gdm starts earlier. Do you have an idea on how to accurately measure the boot time? -- Saludos, Felipe Sateler -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
Bug#581907: sysv-rc: performance regression with CONCURRENCY=makefile
[Felipe Sateler] > I have now stopped bootchart when gnome-panel is loaded. It is not a > perfect hint, but should be closer than the earlier estimates. > Times now seem closer but makefile is still slower. Thank you. I fail to find gnome-panel in bootchart-none-autologin.png. When was it started? I notice Xorg starts after 27 seconds with parallel booting, and 40 seconds with sequencial boot. This make me suspect the regression is really an artifact of how you measure the boot duration, and not a result of the real boot speed slowdown. Happy hacking, -- Petter Reinholdtsen -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
Bug#581907: sysv-rc: performance regression with CONCURRENCY=makefile
On Mon, May 17, 2010 at 01:45, Petter Reinholdtsen wrote: > [Felipe Sateler] >> As per [1], I'm reporting this against the sysv-rc package. >> CONCURRENCY=makefile results in a performance hit to my boot >> sequence, both with and without readahead. > > Thank you. I hope we can find the cause of this slowdown. > > I notice from bootchart-makefile-readahead.png that there are one > startpar process that seem to run longer than expected. It is started > from rcS.d/. This is normally a problem with some daemon failing to > close its file descriptors when it starts. Can you run 'ps -ef|grep > startpar' to see which process this is? These should be reported as > separate bugs. Not sure it affect the boot speed, though. This seems to be timidity. > > I notice netagent sleep for 4 seconds during the boot. Any idea why? I have no idea what net.agent is. > > I also notice X.org seem to start earlier when parallel boot is > enabled. Do you benchmark from the kernel start to the desktop is > ready for use (using autologin) or the login screen is available, or > just to the last init.d script is done? The sequencial boot chart > seem to ignore the time it take to start X, while it is included in > the parallel boot chart. It is up to the login screen. I think bootchart stops when rc2.d is done, so its true the benchmarks could be unfair. I will setup autologin and stop bootchart once the desktop has started. I'm not quite sure how to force a command to fire when gnome has finished loading, will check how to do that. -- Saludos, Felipe Sateler -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
Bug#581907: sysv-rc: performance regression with CONCURRENCY=makefile
[Felipe Sateler] > As per [1], I'm reporting this against the sysv-rc package. > CONCURRENCY=makefile results in a performance hit to my boot > sequence, both with and without readahead. Thank you. I hope we can find the cause of this slowdown. I notice from bootchart-makefile-readahead.png that there are one startpar process that seem to run longer than expected. It is started from rcS.d/. This is normally a problem with some daemon failing to close its file descriptors when it starts. Can you run 'ps -ef|grep startpar' to see which process this is? These should be reported as separate bugs. Not sure it affect the boot speed, though. I notice netagent sleep for 4 seconds during the boot. Any idea why? I also notice X.org seem to start earlier when parallel boot is enabled. Do you benchmark from the kernel start to the desktop is ready for use (using autologin) or the login screen is available, or just to the last init.d script is done? The sequencial boot chart seem to ignore the time it take to start X, while it is included in the parallel boot chart. Happy hacking, -- Petter Reinholdtsen -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org