HELLER@UMASS-CS.CSNET (Number 6) (12/03/85)
We here at UMass/Amherst (COINS Dept.) have a VAXCluster with 11 VAXen (2 780's and 9 750s) and 2 HSC-50s. We also have three graphics devices which all use a DR-11W type interface, that is a fast DMA parallel port off of the UNIBUSS of a particular machine (two devices are presently on one of the 780 and the third is on one of the 750s). Presently, we have some home-grown software to make two of the graphics devices accessable from any of the nodes via DECNet (Ethernet). This is rather slow, partly due to DECNet's overhead and partly due to limited buffer sizes (Ethernet/DECNet packet size limitations). We would like to use the cluster hardware, which is faster than DECNet and allows larger "packets". My idea is to use a psuedo-device with an ACP on each machine, all talking to a detached process running on the machine with the graphics device. The processes would communicate via the cluster hardware and use cluster-wide locks for syncronization. I have written the psuedo-device driver and most of the ACP. At this point I need to figure out how to access the cluster hardware to do bulk data transers. DEC doesn't supply any documentation on how to get at the cluster hardware. They seem to assume that all you will ever do with the cluster is access mass-storage devices (disks and tapes) and have cluster-wide device and batch queues. I have looked at the microfich for CNDRIVER (DEC's DECNet over the CI driver), and have not been able to figure out how the packets get from machine to machine. I would like to know if anyone out there in ARPANet-land has figured out how to get two processes (on two machines) to exchange bulk data (2K to 16K bytes at a time) over the cluster. We are presently running VMS 4.2. Robert Heller VISIONS Group COINS Dept. UMass/Amherst CSNet: Heller@UMass-CS.CSNET ARPANet: Heller%UMass-CS.CSNET@CSNet-Relay.ARPA BITNet: Heller@UMass.BITNET
info-vax@ucbvax.UUCP (12/05/85)
I think the driver you should be looking at is the PADRIVER. I haven't arrived yet with my cluster but soon we will want to be able to access array processors across the cluster and it sounds like what you are doing could be of use. May I ask that you keep me informed of your progress.