| He also observes the following:
From: ISVNET::ISV_SUPPORT "Mail Originating on Digital's Partner Network" 8-JUL-1993 09:38:47.87
To: XLIB::SCHAFER
CC: ISV_SUPPORT
Subj: volz
From: ISVNET::VOLZ "Mail Originating on Digital's Partner Network" 7-JUL-1993 10:37:32.32
To: ISV_SUPPORT
CC: VOLZ
Subj: Ethernet Drivers - Pass on to VMS Engineering?
Hello:
I'd like to suggest that VMS Engineering look at the defaults for the
Ethernet receive ring and receive buffer counts for at least the ISA-0
(EZDRIVER) and SVA-0 (ESDRIVER) controllers.
In particular, we have a VAXcluster running VMS V5.5-2 which has about
55 nodes in it (running over Ethernet) and the System Buffer Unavailable
counts are somewhat high on most of the nodes (especially the main disk
servers). Counts of a few hundred per day are not unusual.
I patched the ESDRIVER to increase the receive buffer count from 9 to
16 and this seems to reduce the System Buffer Unavailable count on
those systems (mostly VAXstation 4000 VLCs). I also patched ETDRIVER
to increase the receive buffer count from 16 to 20, but alas, it
doesn't give the 4 extra buffers to the controller because it only has
16 receive ring entries so the effect seems to be rather minimal (if
any). Note that I did this patching by reviewing the source listings
(so these are not DEC supplied patches).
Are there any other recommendations on trying to reduce the System
Buffer Unavailable counts? (There are no User Buffer Unavailable
counts, so the buffering the various protocols do seems to be
sufficient. We do get occassional send failures - excessive collisions
mostly, but the rate for these is perhaps 10 per day - so the system
buffer unavailable error rate is about a factor of 10 greater and
should be preventable.)
Longer term, we'll probably just have to look at breaking the cluster
into two ... but, it would be nice to avoid that as long as possible.
- Bernie Volz
Process Software Corporation
|
| With regard to the system buffer unavailable problem:
They might want to see if lowering the DECnet executor pipeline quota helps
(see SPEZKO::CLUSTER note 3166). This was covered in an article in the
November 1992 issue of VAXcluster Systems Quorum Journal (available on TIMA).
Also, a retrofitted version of the new V6.0 PEDRIVER (which has congestion
control) has been made available for V5.5-2. I'm told there's a VMSinstallable
kit for this in the save set VAXLAVC02_U2055.A somewhere on the OpenVMS AXP
V1.5 CDROM. A copy is on the net at:
BULOVA::ALPHA$KITS:[V15_SSB.VAX_KIT]VAXLAVC02_U2055.A
(and the kit would also, I'm sure, be available now from the CSCs).
|