vm issues (2)

vm issues (2)

[Dejan Muhamedagic]

Hello,

The new kernel 2.4.21-pre4aa3 is running now, but the box behaves
similarly.  It still swaps quite a lot and much more than the rmap
vm.  Both servers are under the same load.

One difference is the amount of free memory:

 r  b  w   swpd   free   buff  cache  si  so    bi    bo   in cs  us  sy  id
aa:
 0  7  0 5773620 202416 118076 2069716 5330 746  5330   766 4845 5597  12  14  74
rmap:
 0  0  0 3498044  13572   4144 4754596  74   0    75     6  642 598   5   3  92

The aa kernel keeps ~200MB out of 6GB of memory unused.  I'm not
sure, but if we could reduce it perhaps there would be much less
swapping.  Is there a way to achieve this?

Another notable difference between the two vm versions is that the
rmap vm maintains about 80% of memory on the active list and the
aa vm much less: between 4% and 12%.  The rmap vm must use more
CPU, but these servers have a lot of processing power so it is not
noticeable.

Cheers,

Dejan

Re: vm issues (2)

[Andrea Arcangeli]

On Tue, Feb 25, 2003 at 01:13:28PM +0100, Dejan Muhamedagic wrote:
> Hello,
> 
> The new kernel 2.4.21-pre4aa3 is running now, but the box behaves
> similarly.  It still swaps quite a lot and much more than the rmap
> vm.  Both servers are under the same load.
> 
> One difference is the amount of free memory:
> 
>  r  b  w   swpd   free   buff  cache  si  so    bi    bo   in cs  us  sy  id
> aa:
>  0  7  0 5773620 202416 118076 2069716 5330 746  5330   766 4845 5597  12  14  74
> rmap:
>  0  0  0 3498044  13572   4144 4754596  74   0    75     6  642 598   5   3  92
> 
> The aa kernel keeps ~200MB out of 6GB of memory unused.  I'm not
> sure, but if we could reduce it perhaps there would be much less
> swapping.  Is there a way to achieve this?

that is a feature, it guarantees highmem unfreeable allocations like
pagetables can't eat all your normal zone. You can reduce the 200MB with
this boot command:

	lower_zone_reserve=256,256

As to decrease the swapping I just told you how to do that tweaking
vm_mapped_ratio.

> 
> Another notable difference between the two vm versions is that the
> rmap vm maintains about 80% of memory on the active list and the
> aa vm much less: between 4% and 12%.  The rmap vm must use more
> CPU, but these servers have a lot of processing power so it is not
> noticeable.

the theory was that rmap would reduce the cpu utilization but of course
the patch don't do juts rmap.

Andrea

Re: vm issues (2)

[Dejan Muhamedagic]

Andrea:

On Tue, Feb 25, 2003 at 04:32:13PM +0100, Andrea Arcangeli wrote:
> On Tue, Feb 25, 2003 at 01:13:28PM +0100, Dejan Muhamedagic wrote:
> > 
> > The aa kernel keeps ~200MB out of 6GB of memory unused.  I'm not
> > sure, but if we could reduce it perhaps there would be much less
> > swapping.  Is there a way to achieve this?
> 
> that is a feature, it guarantees highmem unfreeable allocations like
> pagetables can't eat all your normal zone. You can reduce the 200MB with
> this boot command:
> 
> 	lower_zone_reserve=256,256

But isn't 200MB too much?  Where would the new setting put the
reserve mark?

> As to decrease the swapping I just told you how to do that tweaking
> vm_mapped_ratio.

Well, it has been set to 500, but it didn't make any difference
(at least no obvious difference).  Is there anything more one
could do about that?  The current level of swapping hurts
performance quite a bit.  This is what the meminfo looks
like:

        total:    used:    free:  shared: buffers:  cached:
Mem:  6341447680 6134513664 206934016        0 66162688 5240639488
Swap: 12582862848 5967990784 6614872064
MemTotal:      6192820 kB
MemFree:        202084 kB
MemShared:           0 kB
Buffers:         64612 kB
Cached:        1982052 kB
SwapCached:    3135760 kB
Active:         304668 kB
Inactive:      4878072 kB
HighTotal:     5358496 kB
HighFree:         2528 kB
LowTotal:       834324 kB
LowFree:        199556 kB
SwapTotal:    12287952 kB
SwapFree:      6459836 kB
BigFree:             0 kB

Cheers!

Dejan

Re: vm issues (2)

[Andrea Arcangeli]

On Tue, Feb 25, 2003 at 05:15:40PM +0100, Dejan Muhamedagic wrote:
> Andrea:
> 
> On Tue, Feb 25, 2003 at 04:32:13PM +0100, Andrea Arcangeli wrote:
> > On Tue, Feb 25, 2003 at 01:13:28PM +0100, Dejan Muhamedagic wrote:
> > > 
> > > The aa kernel keeps ~200MB out of 6GB of memory unused.  I'm not
> > > sure, but if we could reduce it perhaps there would be much less
> > > swapping.  Is there a way to achieve this?
> > 
> > that is a feature, it guarantees highmem unfreeable allocations like
> > pagetables can't eat all your normal zone. You can reduce the 200MB with
> > this boot command:
> > 
> > 	lower_zone_reserve=256,256
> 
> But isn't 200MB too much?  Where would the new setting put the

no, on a 6GB it isn't too much compared to risk wasting several giga of
highmem.

> reserve mark?

the new reserve mark will turn it to around 25mbyte, of course your risk
to run in normal zone shortages increases that way.

> > As to decrease the swapping I just told you how to do that tweaking
> > vm_mapped_ratio.
> 
> Well, it has been set to 500, but it didn't make any difference
> (at least no obvious difference).  Is there anything more one
> could do about that?  The current level of swapping hurts
> performance quite a bit.  This is what the meminfo looks
> like:

you can try 10000.

Andrea