SCHED_ULE makes 256Mbyte i386 unusable

Mon Apr 23 07:49:41 UTC 2018

On 22/4/18 9:43 pm, Rick Macklem wrote:
> Konstantin Belousov wrote:
>> On Sat, Apr 21, 2018 at 11:30:55PM +0000, Rick Macklem wrote:
>>> Konstantin Belousov wrote:
>>>> On Sat, Apr 21, 2018 at 07:21:58PM +0000, Rick Macklem wrote:
>>>>> I decided to start a new thread on current related to SCHED_ULE, since I see
>>>>> more than just performance degradation and on a recent current kernel.
>>>>> (I cc'd a couple of the people discussing performance problems in freebsd-stable
>>>>>   recently under a subject line of "Re: kern.sched.quantum: Creepy, sadistic scheduler".
>>>>>
>>>>> When testing a pNFS server on a single core i386 with 256Mbytes using a Dec. 2017
>>>>> current/head kernel, I would see about a 30% performance degradation (elapsed
>>>>> run time for a kernel build over NFSv4.1) when the server kernel was built with
>>>>> options SCHED_ULE
>>>>> instead of
>>>>> options SCHED_4BSD
> So, now that I have decreased the number of nfsd kernel threads to 32, it works
> with both schedulers and with essentially the same performance. (ie. The 30%
> performance degradation has disappeared.)
>
>>>>> Now, with a kernel from a couple of days ago, the
>>>>> options SCHED_ULE
>>>>> kernel becomes unusable shortly after starting testing.
>>>>> I have seen two variants of this:
>>>>> - Became essentially hung. All I could do was ping the machine from the network.
>>>>> - Reported "vm_thread_new: kstack allocation failed
>>>>>    and then any attempt to do anything gets "No more processes".
>>>> This is strange.  It usually means that you get KVA either exhausted or
>>>> severly fragmented.
>>> Yes. I reduced the number of nfsd threads from 256->32 and the SCHED_ULE
>>> kernel is working ok now. I haven't done enough to compare performance yet.
>>> Maybe I'll post again when I have some numbers.
>>>
>>>> Enter ddb, it should be operational since pings are replied.  Try to see
>>>> where the threads are stuck.
>>> I didn't do this, since reducing the number of kernel threads seems to have fixed
>>> the problem. For the pNFS server, the nfsd threads will spawn additional kernel
>>> threads to do proxies to the mirrored DS servers.
>>>
>>>>> with the only difference being a kernel built with
>>>>> options SCHED_4BSD
>>>>> everything works and performs the same as the Dec 2017 kernel.
>>>>>
>>>>> I can try rolling back through the revisions, but it would be nice if someone
>>>>> could suggest where to start, because it takes a couple of hours to build a
>>>>> kernel on this system.
>>>>>
>>>>> So, something has made things worse for a head/current kernel this winter, rick
>>>> There are at least two potentially relevant changes.
>>>>
>>>> First is r326758 Dec 11 which bumped KSTACK_PAGES on i386 to 4.
>>> I've been running this machine with KSTACK_PAGES=4 for some time, so no change.
> W.r.t. Rodney Grimes comments about this (which didn't end up in this messages
> in the thread):
> I didn't see any instability when using KSTACK_PAGES=4 for this until this cropped
> up and seemed to be scheduler related (but not really, it seems).
> I bumped it to KSTACK_PAGES=4 because I needed that for the pNFS Metadata
> Server code.
>
> Yes, NFS does use quite a bit of kernel stack. Unfortunately, it isn't one big
> item getting allocated on the stack, but many moderate sized ones.
> (A part of it is multiple instances of "struct vattr", some buried in "struct nfsvattr",
>   that NFS needs to use. I don't think these are large enough to justify malloc/free,
>   but it has to use several of them.)
>
> One case I did try fixing was about 6 cases where "struct nfsstate" ended up on
> the stack. I changes the code to malloc/free them and then when testing, to
> my surprise I had a 20% performance hit and shelved the patch.

you might try using uma. especially setting up a non-freeing zone, 
where he system allocates what it needs and then just recycles them.
(man uma)
> Now that I know that the server was running near its limit, I might try this one
> again, to see if the performance hit doesn't occur when the machine has adequate
> memory. If the performance hit goes away, I could commit this, but it wouldn't have that much effect on the kstack usage. (It's interesting how this patch ended
> up related to the issue this thread discussed.)
>
>>>> Second is r332489 Apr 13, which introduced 4/4G KVA/UVA split.
>>> Could this change have resulted in the system being able to allocate fewer
>>> kernel threads/stacks for some reason?
>> Well, it could, as anything can be buggy. But the intent of the change
>> was to give 4G KVA, and it did.
> Righto. No concern here. I suspect the Dec. 2017 kernel was close to the limit
> (see performance issue that went away, noted above) and any change could
> have pushed it across the line, I think.
>
>>>> Consequences of the first one are obvious, it is much harder to find
>>>> the place to map the stack.  Second change, on the other hand, provides
>>>> almost full 4G for KVA and should have mostly compensate for the negative
>>>> effects of the first.
>>>>
>>>> And, I cannot see how changing the scheduler would fix or even affect that
>>>> behaviour.
>>> My hunch is that the system was running near its limit for kernel threads/stacks.
>>> Then, somehow, the timing SCHED_ULE caused resulted in the nfsd trying to get
>>> to a higher peak number of threads and hit the limit.
>>> SCHED_4BSD happened to result in timing such that it stayed just below the
>>> limit and worked.
>>> I can think of a couple of things that might affect this:
>>> 1 - If SCHED_ULE doesn't do the termination of kernel threads as quickly, then
>>>        they wouldn't terminate and release their resources before more new ones
>>>        are spawned.
>> Scheduler has nothing to do with the threads termination.  It might
>> select running threads in a way that causes the undesired pattern to
>> appear which might create some amount of backlog for termination, but
>> I doubt it.
>>
>>> 2 - If SCHED_ULE handles the nfsd threads in a more "bursty" way, then the burst
>>>        could try and spawn more mirror DS worker threads at about the same time.
>>>
>>> Anyhow, thanks for the help, rick
> Have a good day, rick
> _______________________________________________
> freebsd-current at freebsd.org mailing list
> https://lists.freebsd.org/mailman/listinfo/freebsd-current
> To unsubscribe, send any mail to "freebsd-current-unsubscribe at freebsd.org"
>