Date:	Tue, 16 Oct 2007 01:38:57 -0500
From:	Rob Landley <rob@...dley.net>
To:	"Eric W. Biederman" <ebiederm@...ssion.com>
Cc:	Nick Piggin <nickpiggin@...oo.com.au>,
	Theodore Tso <tytso@....edu>,
	James Bottomley <James.Bottomley@...eleye.com>,
	Matthew Wilcox <matthew@....cx>, linux-kernel@...r.kernel.org,
	linux-scsi@...r.kernel.org,
	Suparna Bhattacharya <suparna@...ibm.com>,
	Nick Piggin <piggin@...erone.com.au>
Subject: Re: OOM killer gripe (was Re: What still uses the block layer?)

On Monday 15 October 2007 11:38:33 pm Eric W. Biederman wrote:
> > I don't follow your logic. We don't need SWAP > RAM in order to swap
> > effectively, IMO.
>
> The steady state of a system that is heavily and usably swapping but
> not thrashing is that all of the pages in RAM are in the swap cache,
> at least that used to be the case.

Mind if I throw in some vague and questionable numbers? :)

I vaguely recall that my old 486 laptop with 16 megabyes of ram (circa 1998) 
used to be able to do 3 point something megabytes per second to/from disk, 
according to hdparm -t.  (That was with DMA enabled.)

This means that my old laptop, using sequential writes and not being bogged 
down by excessive seeking, could write its entire memory contents to disk and 
read it back in again in about 10 seconds total (5 write, 5 read).

My current laptop has 2 gigabytes of ram, and hdparm -t /dev/sda says:
  /dev/sda:
   Timing buffered disk reads:  116 MB in  3.01 seconds =  38.54 MB/sec

So that's a little over a factor of 10 speed improvement.  (Although I note 
that I got 30 megabytes/second off of an ATA/100 adapter in 2002, so it's 
barely any faster than it was 5 years ago.)

This means I can expect my current laptop to write out its memory in 50 
seconds (2000/40), and another 50 seconds to read it back in.

So 10 seconds to cycle through memory 10 years ago, vs a little under 2 
minutes today, on systems at roughly the same price point.  And that's 
limited by what the hardware is doing, assuming a _perfect_ linear read/write 
pattern with no seeks.

Oh, and my old 486 had its RAM maxed out.  This one can hold twice as much.  
And heavy seeking sucks more than it used to relative to sequential reads by 
something like a proportional amount (hence the rise of I/O elevators as a 
mitigation strategy), although I haven't got numbers for that handy.

> > I don't know if there is a causal relationship there. I mean, I
> > think it's been a long time since thrashing was ever a viable mode
> > of operation, right?
>
> Right.  But swapping heavily has been a viable mode of operation
> and that the vast gap in disk random IO performance seems to have
> hurt significantly.
>
> It be very clear is used to able to run a problem at little below
> full speed with the disk pegged with swap traffic, and I did this
> regularly when I started out with linux.

The problem is the gap is getting bigger.  The 486-75 laptop mentioned above 
had a 25 mhz 32 bit front side bus.  A quick google suggests my core 2 duo 
has a 667 mhz FSB and I'm guessing a 128 bit data path (two 64-bit channels).

I could boot up memtest86 and get actual benchmarks, but total handwaving for 
a moment, 25*32=800 and 667*128=85376, and the second divided by the first is 
over 100 times as big.  That concurs with the 16mhz->1733 mhz processor speed 
increase.

Factor of 10 disk speed increase, factor of 100 memory speed increase.  Disks 
speeds aren't keeping up with processor and memory increases.  Disk _sizes_ 
are, but speeds aren't.

> > Maybe desktops just have less need for swapping now, so nobody sees
> > it much until something goes _really_ bad. When I'm using my 256MB
> > machine, unused stuff goes to swap.
>
> There is a bit of truth in the fact that there is less need for
> swapping now.  At the same time however swapping simply does not
> work well right now, and I'm not at all certain why.

Do the numbers above help?  It'll only get worse, unless some random new 
technology (maybe http://en.wikipedia.org/wiki/MRAM or something) swoops in 
to change everything, again.

> >> the disk for is very limited.   I wonder if we could figure out
> >> how to push and pull 1M or bigger chunks into and out of swap?
> >
> > Pulling in 1MB pages can really easily end up compounding the
> > thrashing problem unless you're very sure a significant amount
> > of it will be used.
>
> It's a hard call.  The I/O time for 1MB of contiguous disk data
> is about the I/O time of 512 bytes of contiguous disk data.

Hence "the seek sucking even more now" part. :(

I'm sure somebody will eventually write an OLS paper or something on the 
advisability of making swapping decisions with 4k granularity when disks 
really want bigger I/O transactions.  Maybe they already have, somewhere 
between:
http://kernel.org/doc/ols/2007/ols2007v1-pages-53-64.pdf
and
http://kernel.org/doc/ols/2007/ols2007v1-pages-277-284.pdf

Rob
-- 
"One of my most productive days was throwing away 1000 lines of code."
  - Ken Thompson.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists