mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: Ram Pai <linuxram@us.ibm.com>
To: Steven Pratt <slpratt@austin.ibm.com>
Cc: Andrew Morton <akpm@osdl.org>, linux-kernel@vger.kernel.org
Subject: Re: [PATCH/RFC] Simplified Readahead
Date: 05 Oct 2004 10:52:20 -0700	[thread overview]
Message-ID: <1096998740.14033.3.camel@dyn319181.beaverton.ibm.com> (raw)
In-Reply-To: <415DC5D2.8000405@austin.ibm.com>

On Fri, 2004-10-01 at 14:02, Steven Pratt wrote:
> Ram Pai wrote:
> 
> snip...
> 
> >>>>>To summarize you noticed 3 problems:
> >>>>>
> >>>>>1. page cache hits not handled properly.
> >>>>>2. readahead thrashing not accounted.
> >>>>>3. read congestion not accounted.
> >>>>>          
> >>>>>
> >
> >
> >I have enclosed 5 patches that address each of the issues.
> >
> >1 . Code is obtuse and hard to maintain
> >
> >	The best I could do is update the comments to reflect the
> >	current code. Hopefully that should help. 
> >	
> >	attached patch 1_comment.patch takes care of that part to
> >	some extent.
> >
> >
> >2. page cache hits not handled properly.
> >
> >	I fixed this by decrementing the size of the next readahead window
> >	by the number of pages hit in the page cache. Now it slowly
> >	accomodates the page cache hits. 
> >
> >	attached patch 2_cachehits.patch takes care of this issue.
> >
> >3. queue congestion not handled.
> >
> >	The fix is: call force_page_cache_readahead() if we are 
> >	populating pages in the current window.
> >	And call do_page_cache_readahead() if we are populating
> >	pages in the ahead window. However if do_page_cache_readahead()
> >	return with congestion, the readahead window is collapsed back 
> >	to size zero. This will ensure that the next time ahead window
> >	is attempted to populate.
> >
> >	attached patch 3_queuecongestion.patch handles this issue.
> >
> >4. page thrash handled ineffectively.
> >
> >	The fix is: on page thrash detection shutdown readahead.
> >
> >	attached patch 4_pagethrash.patch handles this issue.
> >
> >5. slow read path is too slow.
> >
> >	I could not figure out a way to atleast-read-the-requested-
> >	number-of-pages if readahead is shutdown, without incorporating
> >	the readsize parameter to page_cache_readahead(). So had
> >	to pick some of your code in filemap.c to do that. Thanks!
> >	
> >	attached patch 5_fixedslowread.patch handles this issue.
> >
> >
> >Apart from this you have noticed other issues
> >
> >6.  cache lookup done unneccessrily twice for pagecache_hits.
> >
> >	I have not handled this issue currently. But should be doable
> >	if I introducing a flag, which notes when readahead is
> >	shutdown by pagecahche hits. And hence attempts to lookup
> >	the page only once.
> >	
> >
> >And you have other features in your patch which will be the real
> >differentiating factors.
> >
> >7.  exponential expand and shrink of window sizes.
> >
> >8.  overlapped read of current window and ahead window. 
> >
> >	( I think both are  desirable feature )
> >
> >I did run some premilinary tests using your patch and the above patches
> >and found 
> >
> >your patch was doing slightly better on iozone and sysbench.
> >however the above patch were doing slightly better with DSS workload.
> >  
> >
> 
> Ok, I have re-run the Tiobench tests.  On a single cpu ide based system 
> you new patches have no noticable effect on sequential read performance 
> (a good thing); but on random I/O things went bad :-(.
> 
> Here are the random read results for 16k io with 4GB fileset on 256MB 
> mem, single cpu IDE
> 
>                Stock      w/ patches
> 
>   Threads      MBs/sec      MBs/sec    %diff         diff  
> ---------- ------------ ------------ -------- ------------ 
>          1         1.73         1.72    -0.58        -0.01  
>          4         1.70         1.56    -8.24        -0.14  
>         16         1.66         0.81   -51.20        -0.85  
>         64         1.49         0.68   -54.36        -0.81 
> 
> As you can see somewhere after 4 threads the new patches cause performance to tank.  
> 
> With 512k ios the problem kicks in with less than 4 threads.
> 
>                Stock      w/ patches
>   Threads      MBs/sec      MBs/sec    %diff         diff  
> ---------- ------------ ------------ -------- ------------ 
>          1        18.50        18.55     0.27         0.05 
>          4         8.55         6.59   -22.92        -1.96  
>         16         8.40         5.18   -38.33        -3.22 
>         64         7.34         4.76   -35.15        -2.58 
> 
> 
> Unfortunately this is the _good_ news.  The bad news is that this is much worse on SCSI.
> We lose a few percent on sequential reads for all block sizes and random is just totally screwed.
> 
> Here is the same 16k io requests size with 4GB fileset on 1GB memory on 8way system on single scsi disk.
> 
>                stock        w/ patch
>    Threads      MBs/sec      MBs/sec    %diff         diff   
> ---------- ------------ ------------ -------- ------------ 
>          1         3.43         3.03   -11.66        -0.40   
>          4         4.51         1.06   -76.50        -3.45 
>         16         5.86         1.43   -75.60        -4.43   
>         64         6.13         1.66   -72.92        -4.47 
> 
> 11% degrade even on 1 thread, 75% degrade for 4 threads and above!  This is horribly broken. 
> 
> 
Sorry for the late response. Was out yesterday.

Yes something is broken horribly. Will look into what is broken.  

RP
 


  reply	other threads:[~2004-10-05 17:54 UTC|newest]

Thread overview: 33+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2004-09-23 16:06 Steven Pratt
2004-09-23 22:14 ` Joel Schopp
2004-09-24  0:21 ` Nick Piggin
2004-09-24  2:42 ` Andrew Morton
2004-09-24 15:40   ` Steven Pratt
2004-09-24 16:16     ` Nick Piggin
2004-09-24 16:48       ` Steven Pratt
2004-09-24 22:05     ` Andrew Morton
2004-09-24 22:43       ` Steven Pratt
2004-09-24 23:01         ` Andrew Morton
2004-09-27 15:39           ` Steven Pratt
2004-09-27 19:26             ` Andrew Morton
2004-09-28 10:13               ` Jens Axboe
2004-09-24 22:55       ` Steven Pratt
2004-09-27 20:29         ` Ray Bryant
2004-09-27 21:04           ` Steven Pratt
2004-09-25  0:45       ` Nick Piggin
2004-09-25  1:01 ` Ram Pai
2004-09-25  6:07   ` Ram Pai
2004-09-27 15:30     ` Steven Pratt
2004-09-27 18:42       ` Ram Pai
2004-09-27 20:07         ` Steven Pratt
2004-09-29 18:46           ` Ram Pai
2004-09-29 22:33             ` Steven Pratt
2004-09-29 23:13               ` Andreas Dilger
2004-09-30  2:26                 ` Ram Pai
2004-09-30  5:29                   ` Andrew Morton
2004-09-30 20:20                 ` Stephen C. Tweedie
2004-09-30  1:12               ` Ram Pai
2004-10-01 21:02             ` Steven Pratt
2004-10-05 17:52               ` Ram Pai [this message]
     [not found] <372479081@toto.iv>
2004-09-24  5:00 ` Peter Chubb
2004-09-24 22:57   ` Steven Pratt

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1096998740.14033.3.camel@dyn319181.beaverton.ibm.com \
    --to=linuxram@us.ibm.com \
    --cc=akpm@osdl.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=slpratt@austin.ibm.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®