RE: IPS Issues document

[julian_satran@il.ibm.com]

Hi,

Good job.  I had only an hour to glance over it so I might come later with
some more but here are some observations:

   Performance - has two aspects - related but not entirely equivalent -
   latency and bandwidth. It is obvious that bandwidth is not a
   differentiating factor for any type of interconnect as they use the same
   physical infrastructure in a similar fashion. Latency on the other hand
   is a more difficult factor as it relates to hardware, software and
   protocol. CPU utilization by itself is arguably not a limiting factor -
   as it can be solved by simply adding more CPUs or waiting until the CPUs
   get faster.  All the elements you mention under CPU utilization refer
   mostly to the memory subsystem utilization - and there are no known ways
   to improve on that  as the memory does not get faster as fast as the
   CPUs do (no Moore law for DRAMs!).  Protocol stack processing reduction
   is the mainly improving your "placement" techniques and reducing copy
   operations. Storage subsystems fare better (and so does FCP) since the
   do scatter-gather DMA (and DMA is times faster than copy due to an
   architectural "feature" of DRAMs that use one address setup for several
   memory operations.  This is the reason why we think that giving the
   protocols a chance to do well DMA is so important
   Adoption of the solution by the device vendors is important but not
   critical. Adoption of a "last half-meter" protocol - like the serial-ATA
   pushed by Intel, with all its drawbacks could offer a cheap alternative
   for fiercely competitive environments like desktops with SCSI as a more
   robust and functional cousin. For clusters and RAIDs both can be
   considered and with IP getting adopted by pervasive computing may get IP
   networks to the point at which the price differential will not exist
   anymore
   You do not mention the SBP although, I recall it as the first serial
   SCSI standardized
   On the performance section I think that solution scalability for the
   low-end to the high-end is essential as it is the only way to gain
   "mass" and survive the continuous shift in performance
   Management should include both storage and interconnect.  Any solution
   should be able to use for interconnect management whatever is good for
   network management with only slight concern about the device specifics.
   A factor that I have difficulty expressing is - skills and tools
   availability. By this I mean planers, administrators and (why not!) even
   salesman and their tools. Any solution that we leverage an existing
   skill (and tools) base will have a distinct advantage

Regards,
Julo

Julian Satran - IBM Research at Haifa