Conference iosg::all-in-1_v30

    
    	I don't see why it wouldn't improve performance. I would expect
    	striping to give you more than shadowing, and both to give you more
    	than either one. I suppose it depends on whether that is their
    	bottleneck.
    
    	Maybe a performance expert will give you a better answer.
    
    Regards,
    Paul

...what is disk striping please?

Scott

    
    	Over simplified, you bind a set of disks (say 3) together,
    	then write consecutive blocks on different disks. So block 1 on
    	disk 1, 2 on 2, 3 on 3, 4 on 1 etc... That way you have a logical
    	disk 3 times as big, and you can simultaneously post 3 reads (or
    	writes) when you want 3 blocks and it will (theoretically) complete
    	in a third of the time on average.
    
    	The VAX/VMS striping driver just sits on top of the regular disk
    	drivers and splits the I/Os.
    
    	It has been used for years on Crays with washing machine sized
    	disks, but is gaining popularity with small diameter hard disks.
    
    	Its also one of the RAID levels. WNT has it too. You can also build
    	in redundancy by adding a redundant disk which allows any one
    	disk to fail, and its data can be recreated from the remaining
    	members.
    
    Regards,
    Paul

For striping to be successful, you must evaluate the following:

	1) Is disk I/O limiting system performance?  If not, it will
	   not be beneficial.

	2) Can any I/O problems be resolved through RMS tuning (bucket
	   sizes, private and global buffers, block size)?   This is
	   the first skirmish line in dealing with any I/O performance issues.

	3) If striping is being considered to due to the size of SDAF
	   (to spread SDAF over 2+ disks), it is probably cheaper to buy
	   one larger disk than to use two+ disks, two+ controller
	   channels.

	4) New disk technologies (RF35, RF73, and RA?? coming soon) offer
	   large capacity, really hot performance, at very competitive
	   prices (The customary 'throw hardware at the problem' answer).

	5) The more disks involved in a stripe set, the lower the mean time
	   between failures (more inter-dependent components, more failures).
	   Stripe sets can be shadowed, if you've got the disks or money.

	6) Striping performance is sensitive to the disk geometry, striping
	   chunk size (striping uses chunks, not blocks), and disk controller
	   layout.  The premise is that striping allows multiple parallel
	   I/Os to be directed to multiple members of the stripe set.  In
	   order to be effective, the disks should reside on multiple
	   controller channels. For reads, this may provide lower seek times,
	   and for reads or writes, higher bandwidth.

	7) Striping is most effective when I/O sizes are large.  Most I/Os
	   to interactive time-shared systems are small.  I/Os to SDAF fall
	   into the small category.

In summary, striping is not a magic bullet, but can be useful in solving
I/O or space problems after careful consideration.


Regards,

Marc

    Just to re-inforce what Marc wrote.
    
    In April I worked with Pacificare Health Systems and last week I
    did a follow-up call with them.
    
    They have had no problems (small ones at the begining with MCI mail)
    and have seen no increase in overall response time.  
    
    What they are happy about is "disk space."  They are a large cluster
    with a large number of ALL-IN-1 users.  Heavy mail, etc..
    They had tunned their buckets, and I/O was attribited to spindle
    disk space problems.   We went throuh lots of discussions on what
    disk stripping would give them and the one perception they had going
    in was I/O size.
    
    Tim

Ken Bates' comments:

          <<< COOKIE::DISK$SYSTEM_3:[NOTES$LIBRARY]STRIPING.NOTE;1 >>>
                                     -<  >-
================================================================================
Note 305.6                 Disk STRIPING and ALL-IN-1                     6 of 6
EYLAK::BATES "Ken Bates"                             21 lines   2-OCT-1992 07:19
                             -< Slight correction >-
--------------------------------------------------------------------------------
>	6) Striping performance is sensitive to the disk geometry, striping
>	   chunk size (striping uses chunks, not blocks), and disk controller
>	   layout.  The premise is that striping allows multiple parallel
>	   I/Os to be directed to multiple members of the stripe set.  In
>	   order to be effective, the disks should reside on multiple
>	   controller channels. For reads, this may provide lower seek times,
>	   and for reads or writes, higher bandwidth.

The disks only need to be on different controller channels if the I/O sizes are
large. For "typical" I/O sizes, the contention caused by multiple outstanding
transfers is negligible.

>	7) Striping is most effective when I/O sizes are large.  Most I/Os
>	   to interactive time-shared systems are small.  I/Os to SDAF fall
>	   into the small category.

Actually, the main benefit of striping is load balancing. Load imbalance
usually occurs with applications that have many small requests outstanding.
Data rate may be increased with *very* large I/O's, but this is very atypical.

 - Ken

    
    	I think "Buy faster disks" rather than stripe is good advice. It
    	reminds me of my VMS tuning days where "Buy a faster VAX" was
    	usually better advice than tuning.
    
    Regards,
    Paul