DragonFly BSD
DragonFly users List (threaded) for 2011-04
[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]

Re: Recent concurrency improvements in the AHCI driver and CAM need testing


From: Naoya Sugioka <naoya.sugioka@xxxxxxxxx>
Date: Sun, 10 Apr 2011 23:37:54 -0700

Hello,

This is happened before your recent update, but my laptop showing
CMD=15; timeout
on ahci0.1 when io_acpi is enabled.  This timeout prevents to complete
bootstrap process.
I just wonder this is happened because ahci.0.1 is associated to ATAPI
(DVD-RW) drive without
occupant.

dmesg telles:
ahci0.1: Found ATAPI "TSSTcorp DVD+/-RW TS-U633F D200" serial="R3476GSSA81272"
ahci0.1: tags=0/32 satacap=0202 satafea=0068 NCQ=NO capacity=1.00MB
ahci0.1: f85=0000 f86=0000 f87=4000 WC=notsupp RA=notsupp SEC=notsupp

then start showing a timeout message.

Let me know if you need further information, thank you.
-Naoya

On Sat, Apr 9, 2011 at 9:00 PM, Matthew Dillon
<dillon@apollo.backplane.com> wrote:
>    I've pushed some serious changes to the AHCI SATA driver and CAM.
>
>    One fixes issues where the tags were not being utilized to their fullest
>    extent... well, really they weren't being utilized at all.  I'm not
>    sure how I missed the problem before, but it is fixed now.
>
>    The second ensures that read requests cannot saturate all available
>    tags and cause writes to stall, and vise-versa, and also separates
>    out the read and write BIO streams and treats them as separate entities,
>    which means that reads can continue to be dispatched even if writes
>    saturate the drive's cache and writes can continue to be dispatched
>    even if concurrent read(s) would otherwise eat all available tags.
>
>    The reason the read/write saturation fixes are important is because
>    writes are usually completed instantly since they just go to the drive
>    cache, so even if reads are saturated there's no reason not to push
>    writes to the drive.  Plus when the HD's cache becomes saturated writes
>    no longer complete instantly and would prevent reads from being
>    dispatched if all the tags were used to hold the writes.
>
>    --
>
>    With these fixes I am getting much better numbers with concurrency
>    tests:
>
>    I now get around 37000 IOPS doing random 512-byte sector reads with
>    a Crucial C300 SSD, verses ~8000 or so before the fix.
>
>    And I now get around ~365 IOPS with the same test on a hard drive,
>    verses ~150 IOPS before (remember these are random reads!).
>
>    blogbench also appears to have much better write/read parallelism
>    against the swapcache with the SSD/HD combo.  Memory caches blow
>    out at around blog #1300 on my test boxes.
>
>        With the changes blogbench write performance is maintained through
>        blog #1600 or so, without the changes it drops off at #1300.
>
>        With the changes the swapcache SSD is pushing ~1400 IOPS or so
>        satisfying random read requests.  Without the changes the swapcache
>        SSD is only pushing ~130 IOPS.
>
>        With the changes blogbench is able to maintain a ~60000 article
>        read rate at the end of the test.  Without the changes the
>        read rate is more around ~10000 at the end of the test.  At this
>        stage swapcache has cached a significant chunk of the data
>        in the SSD so the I/O activity is mixed random SSD and HD reads.
>
>    --
>
>    Ok, so I feel a bit sheepish that I missed the fact that the AHCI
>    driver wasn't utilizing its tags properly before.  The difference
>    in performance is phenominal.  Maybe we will start winning some
>    of those I/O benchmark tests now.
>
>                                        -Matt
>                                        Matthew Dillon
>                                        <dillon@backplane.com>
>




[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]