Re: Avoiding deadlock in iSCSI

["Randall R. Stewart" <randall@stewart.chicago.il.us>]

Prasenjit Sarkar/Almaden/IBM wrote:
> 
> Unfortunately, "implementation issues" determine the complexity of a
> protocol and
> consequently, whether the benefits of the protocol are worth the cost.
> 
> BTW, zero-copy TCP/IP stacks have a lot of caveats (e.g. memory alignment
> etc) which
> is why they have never made it to any operating system. There are rough
> implementations
> in Solaris, BSD and Linux but none of them are particularly close to being
> robust
> (I have tried them all).
> 
> What we need is a rational cost-benfit analysis, not simply whether its
> just another
> implementation issue.
> 

Prasenjit:

I have used a zero-copy TCP/IP stack (if I recall correctly) in VxWorks
or was it VRTX... I will grant you these are NOT *NiX systems and 
are just recently beginning to become PosiX compliant... but I
seem to remember the zero-copy symenatics and I don't think there
was a memory alignment requriement.. it has been a while so I
may be mis-remembering.. :)

R

> Prasenjit
> 
>    Prasenjit Sarkar
>    Research Staff Member
>    IBM Almaden Research
>    San Jose
> 
> Michael Krause <krause@cup.hp.com> on 09/17/2000 07:09:51 AM
> 
> To:   Prasenjit Sarkar/Almaden/IBM@IBMUS, Robert Snively
>       <rsnively@Brocade.COM>
> cc:   Kalman Meth/Haifa/IBM@IBMIL, Pierre Labat <pierre_labat@hp.com>,
>       ips@ece.cmu.edu
> Subject:  RE: Avoiding deadlock in iSCSI
> 
> At 09:44 PM 9/16/00 -0700, Prasenjit Sarkar/Almaden/IBM wrote:
> 
> >Actually, because of interrupt coalescing in Gigabit Ethernet adapters,
> >you can have potentially close many SCSI transactions with one interrupt.
> >
> >However, you have hit the nail on the head, parallelism tends to (I'm not
> >saying will always) increase the average number of interrupts per
> >transaction if the parallelism decreases the possibility of interrupt
> >coalescing.
> >
> >And since interrupt coalescing is a statically determined parameter (in
> >current implementations), getting speedups out of parallelism is harder
> >than it appears.
> 
> Interrupt management and parallelism are critical to actual application
> throughput.  There are existing solutions that allow throughput to be
> achieved with intelligent interrupt management.  I do not believe iSCSI
> changes anything along these lines and consider this issue to be
> implementation specific.
> 
> Mike

-- 
Randall R. Stewart
randall@stewart.chicago.il.us or rrs@cisco.com
815-342-5222 (cell) 815-477-2127 (work)