15 respondents commented on the T3E, 9 on PVP systems, and 8 provided suggestions for the next machine.
T3E Queue Related Comments:
4 need T3E queue structure adequate for large-scale problems (longer queues) 2 longer T3E queues for smaller PE amounts (64) 1 T3E turnaround time poor & hard to predict when job will finish 1 too many small jobs on T3E 1 would like to use T3E as a small user 1 T3E scheduler needs improvement: too many idle PEs
Other T3E Comments:
3 T3E praised for great I/O access time, fast communications 1 better T3E I/O (increase number of files open simultaneously) 1 don't like how T3E maps problem onto nodes; most MPPs can take advantage of local communications (nCube, Paragon scale better than T3E) 1 more inodes on the T3E 1 more T3E memory (or better memory management) 1 better 32-bit support for T3E 1 want Cray to continue improving T3E
PVP Comments:
3 want PVPS with more CPU, memory, disk, bandwidth 1 more PVP cycles 1 more interactive PVP CPUs (killeen not enough) 1 better J90 queue structure; management as unit 1 too slow for large (64 mw) PVP jobs; give more weight to large memory jobs 1 need much more memory for C90 class machines; the SV1 should have a lot 1 more disk space on PVPs
Comments about the Next Machine:
2 need at least 1 TB memory for next machine, more memory on T3E or successor 1 worried that performance is meaningless when considering next machine since sustained MPP performance poor compared to peak and likely to get worse 1 more T3E nodes or another parallel machine (mostly would like a much larger T3E) 1 need more, faster computers (needs more CPU speed than memory or I/O) 1 need to do more research on I/O and total memory 1 only need NERSC for parallel computing; do all else on desktop 1 future machine needs CPUs capable of reaching half of peak 1 future machine should take into account mapping of problem onto CPUs