Opened 7 years ago
Last modified 4 years ago
#13697 reopened bug
Laggy performance when SMP and cpuidle are enabled
Reported by: | addos | Owned by: | nobody |
---|---|---|---|
Priority: | normal | Milestone: | Unscheduled |
Component: | Drivers/Power | Version: | R1/Development |
Keywords: | Cc: | ||
Blocked By: | Blocking: | #16125, #16292 | |
Platform: | x86 |
Description
With the last stable release (alpha 4.1), haiku performs fine.
I started testing the nightlies, and deskbar/tracker seem to miss click events. The activity monitor is jerky. It was suggested that I create a video, which I did.
https://drive.google.com/open?id=0BzncnjBGK7NOeDJYTE9vd2FrbW8
From the syslog, this was output:
KERN: int 219, enabled 1, handled 18128, unhandled 15984 KERN: kernel_x86:apic_timer_interrupt__FPv (0x80158820), data 0x00000000, handled 18127
Not sure if that is the problem. I am attaching the syslog from the system.
I started testing different debug options from the bootloader, and the only thing that seemed to make a difference was to disable smp.
Attachments (6)
Change History (56)
by , 7 years ago
comment:1 by , 7 years ago
patch: | 0 → 1 |
---|
comment:2 by , 7 years ago
patch: | 1 → 0 |
---|
comment:3 by , 7 years ago
by , 7 years ago
Attachment: | IMAG0534.jpg added |
---|
comment:7 by , 7 years ago
Could you try with a nightly hrev51974 or newer? (avoid updates because of gcc7 changes)
by , 7 years ago
Attachment: | IMAG0535.jpg added |
---|
comment:15 by , 7 years ago
I disabled the broadcom, ipro1000 driver, hda driver. But none of those disables made it any less choppy. I dropped to kernel_debugger and did ints again with all of those disabled.
by , 7 years ago
Attachment: | IMAG0536-2.jpg added |
---|
comment:16 by , 7 years ago
OK thanks. It seems to be related with SMP then. You could try some older nightlies and see when the regression appeared after hrev45141.
comment:20 by , 6 years ago
Hi, I just tested beta1 on this system, and I am still getting the laggy performance. There are periods of time where performance monitor will stop updating, almost like it is skipping frames or something, and then it will have a burst where it does an update. And there are times where I am mousing over items in the deskbar where the screen doesn't refresh. I am not entirely sure how to debug/troubleshoot what the cause is, though.
comment:21 by , 6 years ago
Can you please retest on a recent nightly? The SMP locking changes may have improved this.
comment:22 by , 6 years ago
Component: | - General → System |
---|
comment:23 by , 6 years ago
Hi, I wanted to say, I just tried nightly build 53159, and the performance does seem significantly better. There are a few pauses, but in general things are much smoother, it is a lot less laggy than it used to be. Now it is at least usable.
comment:24 by , 6 years ago
Actually, maybe I take it back. When I disable multiple cores, and drop it down to a single core. I get a lot more responsiveness out of the system as a whole. So even though it might be a bit better, it still is far from perfect. I tried playing video in webpositive from youtube, and can only really play it decently with only a single core enabled. If I have all the cores enabled, it gets really choppy.
comment:25 by , 6 years ago
Can you recheck against alpha4.1? The ticket description you are talking about missing clicks and the like in Tracker/ActivityMonitor, so knowing if we are at least as good as alpha4 now would be nice.
comment:26 by , 6 years ago
Is there a build of this on the nightly builds downloads page I can download? I didn't seem to see it on there?
comment:28 by , 6 years ago
Hi, I just tested this build on one of the machines I was having problems with, and it does indeed seem to be performing much better. I won't know for certain until I test the other source machine on Monday. But I think this might have made a big difference.
comment:29 by , 6 years ago
I'm confused as to what you are saying. The nightlies are now almost back to alpha4 performance? Or are the even better than alpha4? Or is alpha4 still better?
comment:30 by , 6 years ago
So I have two systems: System 1: the original system, I don't have access to at the moment System 2: the one I have access to at home
When I tried the nightlies on system 1, it was very laggy. Unfortunately, I don't have access to that system until monday.
When I tried alpha 4 on system 2, it seems to be pretty responsive, no lag. I could try to download a recent nightly and try it on system 2, and see if it is laggy again.
comment:31 by , 6 years ago
I am noticing on system 2 with the latest nightly. That activity monitor seems to lag, and refresh periodically, but after long periods of time. I notice, though, that if I have the gl tea pot running, it refreshes faster. What could be causing that? I didn't seem to notice any issues clicking items in the tracker/deskbar with this build.
comment:32 by , 6 years ago
I made the following video of the laggy performance from the latest nightly on system 2.
https://drive.google.com/open?id=19Cm-huWq1Iu76GUSFNiJkEjecnGWO-J3
comment:35 by , 6 years ago
I tried changing various cpu settings in the bios, but none of them had a positive effect. About the only thing that keeps haiku from lagging seems to be booting it with SMP disabled. But performance is subpar when using only a single core. Are there any other suggestions?
comment:36 by , 6 years ago
So I have noticed that if I boot with SMP disabled, I see less lag in general, but if I play audio or video, things are still choppy. The only way I got haiku to actually play audio without interruptions is to disable local apic. What exactly is this option doing differently than disabling SMP? Is there anyway to disable the local apic while still having SMP enabled to see if the OS behaves differently? Is there anything that can be determined about what is going wrong with the options I am disabling and the behavior I am seeing in Haiku?
by , 6 years ago
Attachment: | syslog-20190527.xz added |
---|
comment:37 by , 6 years ago
I uploaded the syslog from system 2 to the ticket. I tested a workaround suggested by Waddlesplash of blacklisting the cpuidle module. This seems to have done the trick on both systems. They both run significantly faster with no lag. I appreciate all the efforts to hunt and track down what the issue seems to have been. My system runs MUCH faster now!
comment:40 by , 6 years ago
I can't find any errata for the Xeon W3520 anywhere about it having broken MWAIT or the like, so I'm not entirely sure why that's the fix.
comment:41 by , 6 years ago
I can't explain it either, I just know that the systems are far less laggy when I blacklist cpuidle.
comment:42 by , 5 years ago
Component: | System → Drivers/Power |
---|---|
Summary: | Laggy performance when SMP is enabled on nightly haiku builds → Laggy performance when SMP and cpuidle are enabled |
comment:44 by , 5 years ago
Hi, I tested this with a newer build, and it seems to be performing well! Thank you for fixing this issue. :)
comment:46 by , 5 years ago
Hi again, I was doing more testing with this, it turns out I was pre-mature. The laggy performance still exists in the latest build. I just wasn't seeing it because I had GL Teapot running. Once I stopped it and started activity monitor again, the performance was laggy again. I was able to blacklist cpuidle and it doesn't lag, so I still don't think this issue is resolved yet. Sorry for prematurely saying it was fixed. :(
comment:47 by , 5 years ago
Resolution: | fixed |
---|---|
Status: | closed → reopened |
comment:49 by , 4 years ago
Blocking: | 16125 added |
---|
comment:50 by , 4 years ago
Blocking: | 16292 added |
---|
Is there any update on this? Is there anything else I can provide to help troubleshoot what is happening? Are there any debug builds I could run that could give you more information?