Re: [PATCH] speed up SATA

From: Eric D. Mudama (edmudama_at_mail.bounceswoosh.org)
Date: 03/28/04

  • Next message: Andrew Morton: "Re: [PATCH] speed up SATA" 
    Date:	Sun, 28 Mar 2004 13:33:57 -0700
To: Nick Piggin <nickpiggin@yahoo.com.au>

    On Sun, Mar 28 at 16:56, Nick Piggin wrote:
    >Eric D. Mudama wrote:
    >>32-MB requests are the best for raw throughput.
    >>
    >>~15ms to land at your target location, then pure 50-60MB/sec for the .5
    >>seconds it takes to finish the operation. (media limited at that point)
    >>
    >>Sure, there's more latency, but I guess that is application dependant.
    >>
    >
    >What about a queue depth of 2, and writing that 32MB in 1MB requests?

    I'm 99% sure that it'll wind up slower... The concept of "ZLR - zero
    latency read?" where they start reading immediately wherever they
    land, then merge that into the host request later, possibly catching
    the front of the command at the very end can help. However, if you
    issue those 32 tags immediately, there's the chance that the drive
    won't recognize the pattern as quickly, and wind up doing additional
    seeks, or even worse, additional, unnecessary dwell... In queueing,
    you could wind up caching data on the back end of 32 host requests
    without the first LBA in cache for any of them, which means you can't
    start those data transfers until you've read a "first LBA", meaning
    you now need to discard something you've intentionally read from the
    media.

    The kicker when handling any request larger than the on-disk cache
    size combined with queueing is that you don't know ahead of time if
    you'll be able to satisfy the request by the time you get there. If
    the drive was already holding megabytes 4-5 in cache, that works
    pretty well as it'll be an instant cache hit, however, if you were
    holding megabytes 4-4.99, there's a chance you will have had to
    discard some of that data to make room for other in-process disk work
    (say you had a buffer almost full of dirty writes and therefore
    insufficient buffer for the entire host request), which means you may
    need to toss megabytes 4.5-4.99 out, in which case you'll need to do
    the disk I/O anyway. In that case, starting at the beginning of the
    entire region and just tossing from cache as you transfer to the host
    works a lot better.

    On a large sequential access, the odds of a cache hit into that
    sequential area is minimal in practice, so we can actually just churn
    the same small chunk of buffer satisfying the large read, while not
    losing any efficiency due to having 6.5MB of dirty random writes ready
    to go on the back end.

    A 32-MB op by definition cannot be atomic with an 8MB DRAM on the
    drive, so once we give up that idea, we optimize our buffer usage for
    maximum cache efficiency and granularity. (Basically, if the drive
    reports a hard error writing a 32MB request, you need to assume the
    entire 32MB request needs to be reissued. You can't assume that just
    because the error was on the last block, that you can "trust" that the
    first LBA of the request actually made it.)

    --eric 

    -- 
Eric D. Mudama
edmudama@mail.bounceswoosh.org
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
  • Next message: Andrew Morton: "Re: [PATCH] speed up SATA"

    Relevant Pages

    • Re: IIS 6.0 Not Caching Images
      ... There is a known bug where if a 204/304 is fulfilled from the kernel ... regardless if it is fulfilled from the kernel response cache or not. ... request is cached in kernel mode (in addition to using WFetch to make the ... server or not -- you need perfmon to determine. ...
      (microsoft.public.inetserver.iis)
    • Re: IIS 6.0 Not Caching Images
      ... which the client would reuse on subsequent requests to the ... IIS is not sending the ETag header on the request correctly. ... > coming out of the kernel mode cache or not. ... There is a known bug where if a 204/304 is fulfilled from the kernel ...
      (microsoft.public.inetserver.iis)
    • Re: Timer question
      ... it seems that the reason you use RemoveCallBack now is just need to ... I think you can consider use a separate background thread to do the ... Cache, you can event not specify a expiration timeout. ... | The request triggered method you recommend is what I am currently doing. ...
      (microsoft.public.dotnet.general)
    • Re: ASP, caching data in the Application object
      ... > Now this works fine as long as there aren't a lot of simultaneous request for this page when the cache has expired. ... > If however the cache has expired and several request are made to the same page, each instance of this page will be taking the task upon him to recreate the cached data, which is unnecessary. ... This code loads the latest data into the currently unused buffer and ...
      (microsoft.public.inetserver.asp.general)
    • Re: Setting session variables other than with set-cookie http header
      ... ISA Server Publishing has a setting to allow caching ... > not get a request for a Set-Cookie, and if it does it will make a note NOT ... It may be able to cache "dynamic" content but there really ... take the load off the DB without trying to cache the page on a proxy. ...
      (microsoft.public.isa)