A5000: no NVRAM?

[Andy Wilcox]

In message <9803192338.AA21487@haven.aegisstar.com>, Phil Brown writes:
>
>For overall bytes/sec throuput, maybe. But what if your goal is to
>complete small, fast-as-possible transactions, that all need to be
>syncd to non-volatile storage before "completion"?
>I don't think you can beat NVRAM, I don't care how fast your rotating
>physical media is. You'll always have a minimum latency in the milliseconds,
>vs microseconds.
>
>Anyone care to argue against that?

Okay  (-;

For overall bytes/sec, _absolutely!

Actually not arguing _against_ it, but consider the effects of
database and/or buffer cache.  Small transactions are likely not 
synchronous to disk, but rather buffered and committed at a 
later date.  That's why we give oracle 1G (or more) of RAM, 
for example.  Going to the disk is vastly slower than memory,
even with the A5000.

If you really want synchronous writes (not a bad idea for a number of
applications, and probably Phil's main point) then the NVRAM is
probably faster, even though it is sitting on the other end of a
slower pipe.  But that is conjecture.  Assuming only rotational
latency on a 10K rpm disk and 8x pipe speed, it might be close.  Even
with fast writes enabled on volumes with small transaction loads, the
fastest writes i've seen reported are .8 - 2 ms (order 1 millisecond),
not order 1 microseconds.

Does anybody have some good numbers on this?  Good topic!

On a related note, a colleague and friend probably on this list, Chuck
Seeger, once remarked that he'd rather have the 1G of cache on a
hardware Raid-5 box closer to the CPU than at the other end of a scsi
chain.


Andy