Conference ssdevo::hsz40_product

852.0. "HSZ50 perf issue?" by NNTPD::"[email protected]" (John McDonald) Wed Apr 23 1997 11:57

I'm working a hot performance issue with a customer. They're running
progress on a 4100, DU4.0b, and an HSZ50 hanging off of a KZPSA. They're
trying to do a database create and it's taking them twice as long as
the original benchmarks did. At first I thought it might be an advfs
issue, but the results of advfsstat showed that advfs wasn't even
breaking a sweat. I then went out to the HSZ50, took one of the
disks out of theur spareset, created a jbod and made the device files
for it on the system. To check the performance, I used 'diskx -p' with
both variable length and fixed length records. here are the results I got
for the variable length test:

Performance test results:

Part-    Transfer  Count of          Read            Write            
Transfer
ition    Size      Transfers         Rate            Rate              Errors
------------------------------------------------------------------------------
-
a         2.0 KB     32768        453.82 KB/sec    342.31 KB/sec           0
a         8.0 KB      8192        881.08 KB/sec    569.72 KB/sec           0
a        14.0 KB      4681          1.27 MB/sec    694.28 KB/sec           0
a        20.0 KB      3276          1.52 MB/sec    889.13 KB/sec           0
a        26.0 KB      2520          1.68 MB/sec    810.71 KB/sec           0
a        32.0 KB      2048          1.82 MB/sec    931.76 KB/sec           0
a        38.0 KB      1724          1.94 MB/sec      1.03 MB/sec           0
a        44.0 KB      1489          1.99 MB/sec    979.28 KB/sec           0
a        50.0 KB      1310          2.08 MB/sec      1.01 MB/sec           0
a        56.0 KB      1170          2.10 MB/sec      1.09 MB/sec           0
a        64.0 KB      1024          2.14 MB/sec      1.12 MB/sec           0
b         2.0 KB     65536        287.79 KB/sec    331.94 KB/sec           0

The maximum observed read rate is     2.14 MB/sec.
The maximum observed write rate is    1.12 MB/sec.

Does anyone have any idea as to why the observed rate are so low?

Thanx.

John McDonald
Atlanta CSC

[Posted by WWW Notes gateway]

p.s.

The writeback cache is enabled and the battery is good.

John

[Posted by WWW Notes gateway]

    What are the facts of the comparison, what was the original 
    configuration that was used as a comparison?
    

    
    I have the same issue at another customer site. Please check
    SMURF::ASE Note 1996.6 and on....

OK,

It's rather late.

But do you happen to know what the MAXIMUM_CACHE_TRANSFER_SIZE of the units
was set to. If it was still set to the default (32 blocks = 16K), then
that may justify your performance problem. If it was set to its maximum
value (as where it should be during write intensive applications, being 1024)
the controller would have used his writeback cache. 

That's my guess

 Are there any hard and fast rules wrt to MAXIMUM_CACHE_TRANSFER_SIZE?
 Any documentation available which explains how to maximize
 hsz performance given various cache sizes, raidsets, policies, etc.

 thanks,
 

    Not  so  much  hard  and  fast, more rule of thumb!! I use the VTDPY 
    tool to work out rough'n'ready details for Average I/O  size.  Then, 
    on  a  unit  by  unit  basis, I set the Cache Xfer Size parameter to 
    make best use of the Read (or WriteBack) cache.

Hi

 On Digital UNIX, I suggest always setting it to 256 (128KB) or larger.