A few days ago I asked:
>I have an RA3000 RAID system with 7 RZ1FC 36 GB disks. 6 of the
>disks are configured as a RAID 5 set and the last one is JBOD.
>The RAID 5 set has 2 partitions with an ADVFS on each. The RA3000
>is connected to an XP-1000 machine. The best sustained performance
>I have seen is lower than 10 MB per second. This is nowhere near
>the 28 MB quoted in the table of specifications in the manual.
>
>I search the archives for RA3000 performance information but failed.
>So, can anybody help me with (pointers to) such performance data?
>Also, hints on how I can improve the performance are of course
>very welcome.
I got very useful replies from
Niels Kokholm <kokholm_at_math.ku.dk>
rich <raf_at_ezunx.com>
Horst Dieter Lenk <lenk_at_mpi-muelheim.mpg.de>
<alan_at_nabeth.cxo.dec.com>
First, I should have written that my simple performance measurements
were done by copying a large file between various file systems while
watching iostat running in another window.
The main problem was that I had not enabled write back cache. With
the write back cache enabled I now get numbers like 18-20 MB/sec.
Comparing this to what the people listed above found makes me feel
quite confident.
Alan gave a lot of interesting information which I would like to
quote below.
Thanks for all the help,
Bjorn
============================
The reply from <alan_at_nabeth.cxo.dec.com> :
The 28 MB/sec capacity of the RA3000 was probably generated
using either a custom seqential read load with a fairly
high queue depth, or a by re-reading the same (relatively
large) cached block of data over and over. Request rate
capacities are generally measured with high queue depth
small size requests that are satisfied by the cache.
If you could describe your I/O load it might offer a clue
what it is load that can expect to get close to the
practical maximum of the controller or not. Some of the
relevant characteristics:
o Read vs. write ratio. This is especially important
for RAID-5 since it has relatively poor write performance
compared to read performance.
o Randomness vs. Sequential I/O. This can affect the RAID-5
performance, since highly sequential I/Os can be combined
in the cache to use RAID-3 write algorithms instead of the
slower RAID-5 write algorithms.
o I/O size. With sufficiently small requests the overhead
of doing the I/O can approach or exceed the time it takes
to move the small amount of data. And, as I recollect,
SCSI command transfers operate at the lowest capability
of the bus; 5 Mhz, 8 bit transfers. This uses up a lot
of bus capacity when the I/O rate is sufficiently high.
With something like iostat, you can estimate the average
I/O size by taking the amount of data transferred between
samples and divide by the number of transfers. Random file
system I/O loads will tend to average around 8 KB. As more
sequential I/O gets added in, the average increases. Raw
disk based applications generate whatever they generate,
though since you're using AdvFS that isn't particularly
relevant. The read vs. write ratio is harder to guess
since you need something that counts them. See if AdvFS
collects file system stats that will count these. If not,
see if the RA3000 management tools will give you access to
VTDPY. If you have CLI access to the RA3000 console you
may be able to use DSTAT. I can send you a DSTAT manual
if you can read PDF files.
Very random I/O loads will be limited by the rate at which
the underlying disks can seek. If you use a worst case guess
that a disk can performance 100 I/Os per second, at an I/O
size of 8 KB, 7 disks can only use 5.6 MB/sec of bandwidth.
For very random I/Os, this guess is probably high for all
but the fastest disks (since seek rate and data size trade
off of each other).
The capacity of any I/O subsystem depends on the character
of the load. Marketing glossies tend to use idealized
loads that show the subsystem to its best advantage. Reality
will tend to lag behind.
===================================================================
Bjorn S. Nilsson Email: Bjorn.S.Nilsson_at_nbi.dk
Niels Bohr Institute or just nilsson_at_nbi.dk
Blegdamsvej 17 Phone: +45 35325283
DK-2100 Copenhagen CellPhone: +45 20701111
Denmark Fax: +45 35325016
Received on Tue Aug 01 2000 - 15:41:43 NZST