My original question:
> We have a 3000/900 system running DU 3.2C.
>
> We have two HSZ40 controllers on separate turbochannel SCSI
> busses. Attatched to each HSZ40, we have at least 1 raid set
> made up of 6 RZ29B-VA drives.
>
> We want to use LSM to stripe accross the controllers to maximize
> throughput on reading the data. Has anybody done the benchmarks
> and is able to tell us the best way to set up the LSM striping?
> Also, is there any advantage/disadvantage to having AdvFS
> overtop of the LSM (We were thinking of having 3 LSM volumes as
> described above combined under AdvFS)?
Thanks to:
alan_at_nabeth.cxo.dec.com
Jim Backer (jbacker_at_cruzio.com)
Mark Zander (mzander_at_spcc.com)
Ohad Imer (ohad_at_comport.com)
The first two responses indicated that LSM has a maximum I/O size
of 64kb and that the best thing to do was local benchmarking.
This is what I did and found that (surprisingly) for the large
I/O's we are doing, peak performance was found with AdvFs overlaid
on LSM. The difference between the UFS and AdvFs was small (less
than 1 Mb/s difference). We also discovered that by striping a
file in AdvFs (with no LSM) and a block size of 256Kb, we got
approximately the same performance.
We also discovered that any improvement above ~6.5 Mb/s was moot
since we are primarily accessing these files from an attatched
MasPar MP-1 which seems to limit the rate of I/O to host mounted
volumes to about that rate.
We chose to use only AdvFs and to stripe any large files which
will be accessed by means other than the MP-1.
Regards,
Mandell Degerness
Received on Wed Dec 11 1996 - 01:53:52 NZDT