I apologize if this e-mail seems a bit disjoint, I'm quite tired from
hauling stuff around today.
I'm not entirely familiar with the system as a whole - but to give a
brief rundown of what I do know:
Context switches, thread prioritization, process statistics keeping,
and access to a handful of other random variables are all serialized
by sched_lock. Process creation, process exit, process scheduling
(schedcpu() access to the allproc_list) are all serialized through the
allproc_lock.
I've discovered that schedcpu()'s serialization needs doesn't fit in
well with sched_lock removal in the presence of a global process list
and global runqueue (I'll skip the tedious details for now). In other
words, I have missing prerequisites. My current plan for this week,
once I get back from Tahoe, is in a separate branch to do the
following:
- replace the global process list with a per-cpu process list hung
off of pcpu protected by a non-interrupt disabling spinlock
pcpu_proclist_lock
- replace the global run queue with a per-cpu runqueue hung off of
pcpu protected by non-interrupt blocking pcpu_runq_lock
Once I have this stable I will integrate it into my branch where I
have replaced sched_lock with per-thread locks and re-do the current
locking I have in choosethread() which I believe causes performance
and stability problems.
At some point it may be desirable to add support for rebalancing the
pcpu process lists to avoid schedcpu/ps/top having to hold the
pcpu_proclist_lock for too long.
Why do I say "non-interrupt blocking?". Currently we have roughly a
half dozen locking primitives. The two that I am familiar with are
blocking and spinning mutexes. The general policy is to use blocking
locks except where a lock is used in interrupts or the scheduler. It
seems to me that in the scheduler interrupts only actually need to be
blocked across cpu_switch. Spin locks obviously have to be used
because a thread cannot very well context switch while its in the
middle of context switching - however, provided td_critnest > 0, there
is no reason that interrupts need to be blocked. Currently sched_lock
is acquired in cpu_hardclock and statclock - so it does need to block
interrupts. There is no reason that these two functions couldn't be
run in ast(). In my tree I set td_flags atomically to avoid the need
to acquire locks when setting or clearing flags. All the timer
interrupt really needs to do for purposes statistics etc. is set a
flag in td_flags indicating to ast() that the current thread is
returning from a timer interrupt so that cpu_hardclock and statclock
are called.
I have more in mind, but I'd like to keep the discussion simple by
focusing on the next week or two.
-Kip
On 6/13/06, Robert Watson <[EMAIL PROTECTED]> wrote:
On Tue, 13 Jun 2006, David Xu wrote:
> On Tuesday 13 June 2006 04:32, Kris Kennaway wrote:
>> On Mon, Jun 12, 2006 at 09:08:12PM +0100, Robert Watson wrote:
>>> On Mon, 12 Jun 2006, Scott Long wrote:
>>>> I run a number of high-load production systems that do a lot of network
>>>> and filesystem activity, all with HZ set to 100. It has also been shown
>>>> in the past that certain things in the network area where not fixed to
>>>> deal with a high HZ value, so it's possible that it's even more
>>>> stable/reliable with an HZ value of 100.
>>>>
>>>> My personal opinion is that HZ should gop back down to 100 in 7-CURRENT
>>>> immediately, and only be incremented back up when/if it's proven to be
>>>> the right thing to do. And, I say that as someone who (errantly) pushed
>>>> for the increase to 1000 several years ago.
>>>
>>> I think it's probably a good idea to do it sooner rather than later. It
>>> may slightly negatively impact some services that rely on frequent timers
>>> to do things like retransmit timing and the like. But I haven't done any
>>> measurements.
>>
>> As you know, but for the benefit of the list, restoring HZ=100 is often an
>> important performance tweak on SMP systems with many CPUs because of all
>> the sched_lock activity from statclock/hardclock, which scales with HZ and
>> NCPUS.
>
> sched_lock is another big bottleneck, since if you 32 CPUs, in theory you
> have 32X context switch speed, but now it still has only 1X speed, and there
> are code abusing sched_lock, the M:N bits dynamically inserts a thread into
> thread list at context switch time, this is a bug, this causes thread list
> in a proc has to be protected by scheduler lock, and delivering a signal to
> process has to hold scheduler lock and find a thread, if the proc has many
> threads, this will introduce long scheduler latency, a proc lock is not
> enough to find a thread, this is a bug, there are other code abusing
> scheduler lock which really can use its own lock.
I've added Kip Macy to the CC, who is working with a patch for Sun4v that
eliminates sched_lock. Maybe he can comment some more on this thread?
Robert N M Watson
Computer Laboratory
Universty of Cambridge
_______________________________________________
freebsd-performance@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "[EMAIL PROTECTED]"
