Pandu E Poluan wrote:
I see.
About "different disks", must this be different physical disks or
different partition is sufficient?
Rgds,
[p]
To give a real world example, I have a server with four cache disks
(500GB SATA), running Squid2.7STABLE6, with two separate disks (250GB
SATA) in a RAID1 for the OS and logs. Load was pretty high with 4 140GB
aufs cache directories (on ext3 mounted noatime, nodiratime) and 200
requests per second (peak). IO wait hovered around 50-60%, with a 5
minute load average around 5. I added a 5GB COSS partition to each
drive*, (so each drive now has a 5GB raw COSS partition, and a 140GB
aufs partition on ext3, mounted noatime, nodiratime) and saw a dramatic
reduction in both IO wait and load (IO wait hovers around 5%, and load
is rarely above .2). Using a COSS file on the same partition as the
aufs directory was not beneficial, nor was using a COSS file on a
separate ext3 partition on the same disk.
For what it's worth, none of this has had much effect on my reported
median service time.
Chris
*cache_dir coss /dev/sdc1 5120 max-size=51200 max-stripe-waste=32768
block-size=4096
