simpler load balance in scheduler
Quoted 6 lines · Paul E. McKenney
Quoted 5 lines · Alex Shi
Hi Paul,
Would you like to give some comments when you'r free? Especially for the 3rd point.
Quoted 7 lines
Quoted 1 line
Appreciate for your any input!
Hello, Alex,
I guess my main question is why we are discussing this internally rather than on LKML. Is this to be some writeup somewhere? (That would be a good thing.)
Thanks Paul!
I very appreciate your great comments and suggestions!
I thought it could be better to collect more comments before send to Peter and LKML. But since the main idea has no much change, and Paul had give very valuable suggestions. Rewrite/resend may cover the Paul's wise comments. So, I'd forward this email to LKML. sorry if the format doesn't looks good.
Quoted 18 lines
On to #3 below...
Quoted 15 lines · Alex Shi
Quoted 14 lines
Paul & Vincent & Morten,
The following rough idea get during this KS. I want to have internal review before send to LKML. Would you like to give some comments?
==========================
1, Current scheduler load balance is bottom-up mode, each CPU need initiate the balance by self. Like in a integrate computer system, it has smt/core/cpu/numa, 4 level scheduler domains.
If there is just 2 tasks in whole system that both running on cpu0. Current load balance need to pull task to another smt in smt domain, then pull task to another core, then pull task to another cpu, finally pull task to another numa. Totally it is need 4 times task moving to get system balance.
As Vincent pointed out, idle balance may help a little for this task moving. But it depends on the CPU topology and which cpu is just idle in chance, so can not help much.
Quoted 31 lines
Quoted 31 lines
Quoted 31 lines
Generally, the task moving complexity is
O(nm log n), n :=nr_cpus, m :=nr_tasksPeterZ has a excellent summary and explanation for this in
kernel/sched/fair.c:4605Another weakness of current LB is that every cpu need to get the other cpus' load info repeatedly and try to figure out busiest sched group/queue on every sched domain level. but may not conduct a task moving, one of reasons is that cpu can only pull task, not pushing.
2, Consider huge cost of task moving: CS, tlb/cache refill, and the useless remote cpu load info getting. If we can have better solution for load balance, like reduce the balance times to.
O(m) m := nr_tasksIt will be a great win on performance. like above example, we can move task from cpu0 direct to another numa. that only need 1 task moving, save 3 CS and tlb/cache refill.
To get this point, a rough idea is changing the load balance behaviour to top-down mode. Say let each of cpu report self load status on per-cpu memory. And a baby-sitter in system to collect these cpus load info, then decide how to move task centralize, finally send IPI to each hungry cpu to let them pull load quota from appointed cpus.
Like in above example, the baby-sitter will fetch each cpus' load info, then send a pull task IPI to let a cpu in another numa pull one task from cpu0. So in the task pulling, we still just involved 2 cpus, can reuse
move_tasksfunctions.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
BTW, the baby-sitter can care all kind of balance, regular balance, idle balance, wake up balance.
3, One of concern of top-down mode is that baby-sitter need remote
The new mode is more like a flat mode balance. But if numa access is really unbearable, we may give a baby-sitter for each of NUMA node, then reduce the numa balance frequency, and do more frequency internal numa node balance. -- I hope that's not needed.
Quoted 23 lines
Quoted 5 lines
Quoted 5 lines
access cpu load info on top domain level every times. But the fact is current load balance also need to get remote cpu load info for top level domain balance. and more worse, such remote accessing maybe useless. -- since there is just one thread reading the info, no competitor writer, Paul, do you think it is worthy concern?
Might be -- key thing is to measure it on a big system. If it is a problem, here are some workarounds with varying degrees of sanity:
1. Reduce cache misses by keeping three values:
a. Current load value.
b. Last exported value. (Should be in same cache line
as a.)
c. Exported value. (Should be in separate cache line.)The avoid writing to c unless a has deviated sufficiently from a. If the load values stay constant for long time periods, this can reduce the number of cache misses. On the other hand, it introduces an extra compare and branch -- though this should not be a problem if the load value is computed relatively infrequently.
Quoted 6 lines
2. As above, but provide additional information to allow the
other CPU to extrapolate values. For example, if a CPU goes idle, some of its values will decay exponentially. The remote CPU could compute the decay rate, removing the need for the subject CPU to wake up to recompute its value. (Maybe you already do this?)
The idle_cpus_mask should be one of additional info to allow extrapolation action. As Vincent explained, this variable set when CPU enter idle and cleared during next tick. So we mayn't need to look at the cpu load for idle cpu.
And yes, current scheduler has similar behaviour.
Quoted 7 lines
Similarly if a CPU remains CPU-bound with given runqueue loading for some time.
3. Allow the exported values to become inaccurate, and resample
the actual values remotely if extrapolated values indicate that action is warranted.
It is a very heuristic idea! Could you give a bit more hints/clues to get remote cpu load by extrapolated value? I know RCU use this way wonderfully. but still no much idea to get live cpu load...
Quoted 8 lines
There are probably other approaches. I am being quite general here because I don't have the full picture of the scheduler statistics in my head. It is likely possible to obtain a much better approach by considering the scheduler's specifics.
Quoted 2 lines
Quoted 2 lines
BTW, to reduce unnecessary remote info fetching, we can use current
idle_cpus_maskin nohz, we just skip the idle cpu in this cpumask simply.
[..]
Quoted 6 lines
Thanx, Paul
Quoted 3 lines
Quoted 3 lines
4, From power saving POV, top-down give the whole system cpu topology info directly. So beside the CS reducing, it can reduce the idle cpu interfere by a transition task. and let idle cpu sleep better.
Signature · 2 lines
Thanks
AlexQuoted 2 lines · Alex Shi
Quoted 1 line · Paul E. McKenney
On Fri, Dec 13, 2013 at 06:09:47PM +0800, Alex Shi wrote:
[ . . . ]
Quoted 7 lines
Quoted 3 lines
3. Allow the exported values to become inaccurate, and resample
the actual values remotely if extrapolated values indicate that action is warranted.It is a very heuristic idea! Could you give a bit more hints/clues to get remote cpu load by extrapolated value? I know RCU use this way wonderfully. but still no much idea to get live cpu load...
Well, as long as the CPU continues doing the same thing, for example, being idle or running a user-mode task, the extrapolation should be exact, right? The load value was X the last time the CPU changed state, and T time has passed since then, so you can calculated it exactly.
The exact method for detecting inaccuracies depends on how and where you are calculating the load values. If you are calculating them on each state change (as is done for some values for NO_HZ_FULL), then you simply need sufficient synchronization for geting a consistent snapshot of several values. One easy way to do this is via a per-CPU seqlock. The state-change code write-acquires the seqlock, while those doing extrapolation read-acquire it and retry if changes occur. This can have problems if too many values are required and if changes occur too fast, but such problems can be addressed should they occur.
Does that help?
Thanx, Paul
Quoted 20 lines
Quoted 7 lines
There are probably other approaches. I am being quite general here because I don't have the full picture of the scheduler statistics in my head. It is likely possible to obtain a much better approach by considering the scheduler's specifics.
Quoted 2 lines
Quoted 2 lines
BTW, to reduce unnecessary remote info fetching, we can use current
idle_cpus_maskin nohz, we just skip the idle cpu in this cpumask simply.[..]
Quoted 6 lines
Thanx, Paul
Quoted 3 lines
Quoted 3 lines
4, From power saving POV, top-down give the whole system cpu topology info directly. So beside the CS reducing, it can reduce the idle cpu interfere by a transition task. and let idle cpu sleep better.
Signature · 2 lines
Thanks Alex
Quoted 10 lines · Paul E. McKenney
Quoted 2 lines · Alex Shi
Quoted 1 line · Paul E. McKenney
On Fri, Dec 13, 2013 at 06:09:47PM +0800, Alex Shi wrote:
[ . . . ]
Quoted 4 lines
Quoted 3 lines
3. Allow the exported values to become inaccurate, and resample
the actual values remotely if extrapolated values indicate that action is warranted.
Quoted 8 lines
Quoted 3 lines
It is a very heuristic idea! Could you give a bit more hints/clues to get remote cpu load by extrapolated value? I know RCU use this way wonderfully. but still no much idea to get live cpu load...
Well, as long as the CPU continues doing the same thing, for example, being idle or running a user-mode task, the extrapolation should be exact, right? The load value was X the last time the CPU changed state, and T time has passed since then, so you can calculated it exactly.
It's a good idea that I never thought before. Thanks a lot!
Quoted 10 lines
The exact method for detecting inaccuracies depends on how and where you are calculating the load values. If you are calculating them on each state change (as is done for some values for
NO_HZ_FULL), then you simply need sufficient synchronization for geting a consistent snapshot of several values. One easy way to do this is via a per-CPU seqlock. The state-change code write-acquires the seqlock, while those doing extrapolation read-acquire it and retry if changes occur. This can have problems if too many values are required and if changes occur too fast, but such problems can be addressed should they occur.
I thought about the seqlock, but it is clearly not scalable. Anyway, load balance don't be very accurate, so maybe atomic operate for exported per cpu load in balance is acceptable.
Quoted 2 lines
Does that help?
Yes, very helpful! :)
Quoted 24 lines
Thanx, Paul
Quoted 20 lines
Quoted 7 lines
There are probably other approaches. I am being quite general here because I don't have the full picture of the scheduler statistics in my head. It is likely possible to obtain a much better approach by considering the scheduler's specifics.
Quoted 2 lines
Quoted 2 lines
BTW, to reduce unnecessary remote info fetching, we can use current
idle_cpus_maskin nohz, we just skip the idle cpu in this cpumask simply.[..]
Quoted 6 lines
Thanx, Paul
Quoted 3 lines
Quoted 3 lines
4, From power saving POV, top-down give the whole system cpu topology info directly. So beside the CS reducing, it can reduce the idle cpu interfere by a transition task. and let idle cpu sleep better.
Signature · 2 lines
Thanks Alex
Signature · 2 lines
Thanks
Alex