The following user messages are collected from the HIPPI Reflector
From: owner-hippi-ext
Subject: Sbus Adapter
Date: Thursday, 13 October. 1994 16:44
I've been using the Chi Systems (1-510-440-0400) Sbus HIPPI board for the past 8-10 months or so on a Sparc 10 running SunOS 4.1.3. Talking between two Sparc 10s over HIPPI, using the 'ttcp' benchmark, I get around ~5.0MB/s (what I would expect-limited by the Sun). I should be getting Sparc 20s in soon. Talking between the Sun and the old SGI, running IRIX 4.0.5, with a Chi HIPPI board in it, I get ~3.0MB/s (This is definitely due to the SGI I have). These are all using small window sizes. I haven't tested against IRIX 5.2 as of yet. If anyone out there can tell me how to do large windows on the Sun, I'd love to hear from you.
Don't ask me about the Sbus Adapter to the CM5 HIPPI. The CM5 doesn't like the small windows.
jjm
We have also been using the Chi SBus HIPPI card, only we have been using it on a Sparc 10/41 running Solaris 2.3. The TCP performance is about the same as Jim reports. I've collected performance numbers for raw HIPPI transfers from the SS 10 to a HIPPI tester. We typically see 18 to 20 MB/sec in this configuration. Chi reports that an SS 10 running SunOS 4.1.3 can turn about 24 MB/sec in the same configuration (will Solaris ever be as fast as SunOS 4.1.3?).
We have also developed an IPI-3 pseudo driver which sits atop the Chi raw driver. First party IPI-3 transfers to a MaxStrat Gen4L disk array yielded 10 MB/sec for reads and 12 MB/sec for writes. Third party transfers have shown numbers that range from 12 MB/sec to 16 MB/sec (the timing of the 3rd party transfers still needs some work and may not be entirely accurate). Achieving these numbers with IPI-3 requires using fairly large buffer (2 MBytes) to amortize the cost of control.
GENROCO also manufactures an SBus/HIPPI card, but the expected performance will probably not be significantly different than the Chi card. The real bottleneck, as I understand it, is the M2S chip which introduces substantial latency into SBus transfers. GENROCO is also building PCI, TurboChannel, and FutureBus versions of their HIPPI card.
For more info on the Chi card, contact Barry Gleaton < barry@chisys.com>
,
for info on the GENROCO card, contact Carl Pick < carl@genroco.com>
.
Regarding the question of large windows on the Sun, you can increase window
size with ttcp using the '-b' option, which sets socket buffer size. A socketbuffer of 64K translates to a window of 64K. If you want a larger window, you'll need the RFC1323 extensions. I believe Bob Braden may have diffs for these extensions for SunOS 4.1.3.
mb
Mark Boolootian booloo@llnl.gov +1 510 423 1948
Using 'txtime' and 'rxtime' programs bundled with CHI's cards, I am measuring the performance of memory to memory data transfers via HiPPI between two SparcServer 1000's. The SparcServers are configured with CHI SBus HiPPI interface cards and are using the SBus-HiPPI raw character drivers. The workstations are interconnected using NSC's ES-1 switch.
The CHI driver uses Sun's bit copy (hwbc_copy()) routine which makes use of the system cache. Unfortunately, a hardware bug on the Sparc 1000 (Sun4d architecture) makes using Sun's version of the bit copy routine improbable until a work around is developed. Because of that, we are using CHI's software version of this routine which is giving us poor performance results (peak of 19 Mbits/sec).
Drew Firment
Hughes Information Technology Corp.
This is one of the CERN High Speed Interconnect pages - 12 Juny 1995 - Arie van Praag