Summary:LSM Advice Please

[Scott Horton <scotthorton7@xxxxxxxxxxx>]

Thanks very much to all that responded!!

idea:
The default stripe size with LSM should be fine, only issue might be if
you
do lots of writes - a non write-back-cache write can take up to 70ms,
and it
might be in your benefit to add a the LVD raid card - KZPCC, which will
increase performance and capacity through raid 5, which is not bad if
you
have WBC enabled.

idea:
>From my 331 book:  (Designing and Implementing Compaq SAN solutions)

i.e. This is for HSG controllers, but I assume it should apply to LSM.

You have 2 basic scenarios - High Request Rate and High Data Rate.

(With lots of small files, I assume you are in a High Request Rate
situation.)

Again from the book:

Stripe Width = Chunk Size = Amount of data on a single disk/stripe

For High Request Rate:

High Locality: Chunk Size = 10x Average Transfer Size
Low Locality: Chunk Size = 20x Average Transfer Size
Unknown Locality: Chunk Size = 15x Average Transfer Size

Chunk size is specified in sectors, 512 bytes/sector.

Chunk size should be a prime number to "reduce the sequential response
time variance".  (I don't know what this means.)

For High Data Rate  (i.e. Large Sequential i/o requests):

Chunk Size = 17 is the recommended value.


>>  Hello all,
 >>  I need advice on the most practical way to configure a stripe set
to
 >>  achieve
 >>  the greatest  R/W performance.  I have an ES45 with the extra disk
cage,
 >>  and
 >>  an external StorageWorks dual bus ultra3LVD box.  Both controllers
sit
 >>  on 66MHZ
 >>  internal buses. Now I realize that I should stripe across each bus,
but
 >>  I'm not sure
 >>  how wide the stripe width should be. The data being written is very

 >>  small..ie 1.5K to
 >>  10K. So this would be many small files. The default stripe width is
64k,
 >>  so it seems
 >>  that I would be losing most of the striping benefit unless I
lowered the
 >>  stripe width.
 >>  How low can (or should) I go with the width before the sheer
overhead of
 >>  managing
 >>  the set slows things down.  ANY ideas are welcomed.

 >>  Thanks,
 >>  Scott