Re: [RFC][PATCH] mm: the page of MIGRATE_RESERVE don't insert into pcp

From: KOSAKI Motohiro
Date: Thu Nov 13 2008 - 23:31:21 EST


Hi

Sorry for late responce.
Honestly, I have many bisiness trip in this month ;-)


> > > > > What your patch may help is the situation where the system is under intense
> > > > > memory pressure, is dipping routinely into the lowmem reserves and mixing
> > > > > with high-order atomic allocations. This seems a bit extreme.
> > > >
> > > > not so extreame.
> > > >
> > > > The linux page reclaim can't process in interrupt context.
> > > > Sl network subsystem and driver often use MIGRATE_RESERVE memory although
> > > > system have many reclaimable memory.
> > > >
> > >
> > > Why are they often using MIGRATE_RESERVE, have you confirmed that? For that
> > > to be happening, it implies that either memory is under intense pressure and
> > > free pages are often below watermarks due to interrupt contexts or they are
> > > frequently allocating high-order pages in interrupt context. Normal order-0
> > > allocations should be getting satisified from elsewhere as if the free page
> > > counts are low, they would be direct reclaiming and that will likely be
> > > outside of the MIGRATE_RESERVE areas.
> >
> > if inserting printk() in MIGRATE_RESERVE, I can observe MIGRATE_RESERVE
> > page alloc easily although heavy workload don't run.
> > but, there aren't my point.
> >
>
> That's interesting. What is the size of a pageblock on your system and
> is min_free_kbytes aligned to that value? If it's not aligned, it would
> explain why MIGRATE_RESERVE pages are being used before the watermarks
> are hit.

hmm, I don't have it yet.
ok, I should investigate more.


> > ok, I guess my patch description was too poor (and a bit pointless).
> > So, I retry it.
> >
> > (1) in general principal, the system should effort to avoid oom rather than
> > performance if memory shortage happend.
> > MIGRATE_RESERVE directly indicate memory shortage happend.
> > and pcp caching can prevent another cpu allocation.
>
> MIGRATE_RESERVE does not directly indicate a memory shortage has
> occured. Bear in mind that a number of pageblocks are marked
> MIGRATE_RESERVE based on the value of the watermarks. In general, the
> minimum number of pages kept free will be in the MIGRATE_RESERVE blocks
> but it is not mandatory.
>
> > (2) MIGRATE_RESERVE is never searched by buffered_rmqueue() because
> > allocflags_to_migratetype() never return MIGRATE_RESERVE.
> > it doesn't work as cache.
> > IOW, it don't help to increase performance.
>
> This is true. If MIGRATE_RESERVE pages are routinely being used and placed
> on the pcp lists, the lists are not being used to their full potential
> and your patch would make sense.
>
> > (3) if the system pass MIGRATE_RESERVE to free_hot_cold_page() continously,
> > pcp queueing can reduce the number of grabing zone->lock.
> > However, it is rate. because MIGRATE_RESERVE is emergency memory,
>
> Again, MIGRATE_RESERVE is not emergency memory.
>
> > and it is often used interupt context processing.
> > continuous emergency memory allocation in interrupt context isn't so sane.
> >
> > Then, unqueueing MIGRATE_RESERVE page doesn't cause performance degression
> > and, it can (a bit) increase realibility and I think merit is much over demerit.
> >
>
> I'm now inclined to agree if you have shown that MIGRATE_RESERVE pages are
> routinely ending up on the PCP lists.

Thanks!

So, now, I have two todo issue.
- I should mesure performance.
- I should investigate why MIGRATE_RESERVE is used on my machine.

I expect I finish to the end of next week.

> > Yup, I believe at that time your decision is right.
> > However, I think the condision was changed (or to be able to change).
> >
> > (1) legacy pcp implementation deeply relate to struct zone size.
> > and, to blow up struct zone size cause performance degression
> > because cache miss increasing.
> > However, it solved cristoph's cpu-alloc patch
>
> Indeed.
>
> > (2) legacy pcp doesn't have total number of pages restriction.
> > So, increasing lists directly cause number of pages in pcp.
> > it can cause oom problem on large numa environment.
> > However, I think we can implement total number of pages restriction.
> >
>
> Yes although knowing what the right size for each of the lists should be
> so that the overall PCP lists are not huge is a tricky one.

Thank you for good advice.




--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/