Re: [squid-users] Squid high bandwidth IO issue (ramdisk SSD)

Generally large amounts of CPU being spent in IO wait means that the
driver is not well-written or the hardware requires extra upkeep to
handle IO operations.

What hardware in particular are you using?

This was one of those big differences between IDE and SATA in the past
btw. At least under Linux in the distant past, a lot of the IDE
drivers would have to manually transfer the data using PIO rather than
having a bus-master DMA transfer occur like many SCSI cards did. This
was counted to "IO wait."

Investigate what your storage driver is doing. :)

HTH,

Adrian

2009/8/2 smaugadi <adi_at_binat.net.il>:
>
> Well I'm seeing that the CPU is taking a lot of time waiting for outstanding
> disk I/O request.
> Adi
>
> Adrian Chadd-3 wrote:
>>
>> Are you seeing high IO wait CPU use, or high IO wait times on IO?
>>
>>
>>
>> Adrian
>>
>> 2009/8/2 smaugadi <adi_at_binat.net.il>:
>>>
>>> Dear Adrian,
>>> Well my conclusion that this is an IO problem came from the fact that I
>>> see
>>> huge IO waits as the volume of traffic increase (with tools such as
>>> mpstat),
>>> when using ramdisk there is no such issue.
>>> I have configured the SSD drive with ext2, no journal, noatime. Used the
>>> “noop” I/O scheduler.
>>> In /etc/fstab
>>> /dev/sdb1               /cache                  ext2 defaults,noatime 1 2
>>>
>>> hdparm results:
>>> hdparm -t /dev/sdb1
>>>
>>> /dev/sdb1:
>>>  Timing buffered disk reads:  304 MB in  3.01 seconds = 100.93 MB/sec
>>> ----
>>> hdparm -T /dev/sdb1
>>>
>>> /dev/sdb1:
>>>  Timing cached reads:   4192 MB in  2.00 seconds = 2096.58 MB/sec
>>>
>>> Any ideas?
>>>
>>> Regards.
>>>
>>>
>>>
>>> Adrian Chadd-3 wrote:
>>>>
>>>> 2009/8/2 smaugadi <adi_at_binat.net.il>:
>>>>>
>>>>> Dear Adrian,
>>>>> During the implementation we encountered issues with all kind of
>>>>> variables
>>>>> such as:
>>>>> Limit of file descriptors (now the squid is using 204800).
>>>>> TCP port range was low (increased to 1024 65535) TCP timers (changed
>>>>> them)
>>>>> The ip_conntrack and hash size were low (now 524288 262144
>>>>> respectively)
>>>>>
>>>>> Now we are at a point that IO is the only issue.
>>>>
>>>> What profiling have you done to support that? For example, one of the
>>>> issues I had which looked like IO performance was actually because the
>>>> controller was completely unhappy. Upgrading the firmware on the
>>>> controller card signficantly increased performance.
>>>>
>>>> But I think you need to post some further information about the
>>>> problem. "IO" can be rooted in a lot of issues. :)
>>>>
>>>>
>>>> Adrian
>>>>
>>>>
>>>
>>> --
>>> View this message in context:
>>> http://www.nabble.com/Squid-high-bandwidth-IO-issue-%28ramdisk-SSD%29-tp24775448p24776193.html
>>> Sent from the Squid - Users mailing list archive at Nabble.com.
>>>
>>>
>>
>>
>
> --
> View this message in context: http://www.nabble.com/Squid-high-bandwidth-IO-issue-%28ramdisk-SSD%29-tp24775448p24776478.html
> Sent from the Squid - Users mailing list archive at Nabble.com.
>
>
Received on Sun Aug 02 2009 - 10:27:40 MDT