[Disksim-users] Synthetic sequential read performance
Anjo Vahldiek
vahldiek at mpi-sws.org
Thu Sep 8 09:37:04 EDT 2011
Dear DiskSim users
Currently I'm facing an issue for the clean disksim 4.0 distribution
using the Seagate Cheetah model provided on the CMU website in which
sequential reads take longer than local reads. I enclosed the used
parameter file. (changed the controller to type 3, type 1 shows the same
result)
I tried to locate the problem using a synthetic seq. read workload and
it seems that the following happens:
1. read access (starts at blkno)
- seek to cylinder
- rotate until first read block
- start reading and transfering
- finish request
2. read access (starts at blkno + 8)
- misses the block blkno + 8 (first block)
- reads the following 7
- make one full rotation to read blkno + 8
3. read access (starts at blkno + 16)
- rotate to blkno + 16
- read 8 blocks and finish request
4. read access (starts at blkno + 24)
- misses block blkno + 24 (first block)
- same behaviour as 2. read access
Basically every second access introduces a full rotation, because it
misses the first block. Unfortunately it does not read during the
rotation into a disk buffer segment such that subsequent accesses would
return the result immediately.
The avg. response time per request is about 4ms which results in a
throughput of about 1MB/s. Usually I would expect for a 15k rpm disk
70-80MB/s.
Hopefully one of you has seen this before and may be able to tell me how
I have to configure or change the disk model parameters. Any idea would
be very helpful.
Thank you very much in advance.
Anjo Vahldiek
-----
Max Planck Institute for Software Systems (MPI-SWS)
PhD-Student
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: Seagate-Cheetah15k5.parv_base
URL: <http://lists.andrew.cmu.edu/pipermail/disksim-users/attachments/20110908/9f07afe6/attachment.ksh>
More information about the Disksim-users
mailing list