pase@ogcvax.UUCP (09/07/87)
It occured to me that my previous postings on iPSC communication times were incomplete without times for global broadcasts. The iPSC has the ability to broadcast messages as a single operation to different sub-cubes up to the size of the whole cube. The times posted in this article are only for global broadcasts (i.e. broadcasts to the whole cube). Each time is in seconds, and is the average of 10 broadcasts. Each broadcast was initiated by node zero, and was echoed by node 2^d-1 (`d' is the cube dimension). Times for smaller sub-cubes may be treated in a future posting should there be any interest. BTW, I'm compiling a list of functions which divulge information useful to hypercube users. Examples of this are routing algorithms, algorithms which estimate message transmission time, determine the distance between two nodes, the number and id's of nodes at a certain distance, the number of states needed by a Markov chain to model a hypercube, etc. Some routines apply specifically to the iPSC (e.g. message transmission times), others do not. Currently the list of functions is quite small but growing. If anyone is interested in having their routines added to this list, I would appreciate seeing the code. I am maintaining this library in `C', but submissions may be in any reasonable language. Please include the author of the algorithm, if possible, so that proper credit may be given. If anyone would like copies of what I have that is fine also. Send submissions and requests to: ...verdix!ogcvax!pase ...ucbvax!tektronix!ogcvax!pase pase@Oregon-Grad.csnet Now for the timings I promised... (Note the anomalous spikes on the d3 broadcast with a message of 7.5K and 8.0K bytes, and d5 with 13.5K and 14.0K bytes. As best I am able to tell, those spikes are really there. They appeared in every test, even when tested for alone.) Cube Dimension Message Size 1 2 3 4 5 ---------------------------------------------- 0.0K 0.0012 0.0017 0.0022 0.0027 0.0035 0.5K 0.0032 0.0027 0.0052 0.0047 0.0055 1.0K 0.0035 0.0047 0.0050 0.0065 0.0077 1.5K 0.0037 0.0065 0.0092 0.0125 0.0162 2.0K 0.0055 0.0072 0.0110 0.0150 0.0195 2.5K 0.0060 0.0110 0.0152 0.0185 0.0252 3.0K 0.0075 0.0117 0.0170 0.0225 0.0282 3.5K 0.0077 0.0150 0.0207 0.0255 0.0315 4.0K 0.0100 0.0160 0.0222 0.0295 0.0370 4.5K 0.0100 0.0180 0.0262 0.0347 0.0432 5.0K 0.0115 0.0192 0.0275 0.0345 0.0455 5.5K 0.0120 0.0222 0.0315 0.0410 0.0520 6.0K 0.0125 0.0227 0.0330 0.0432 0.0542 6.5K 0.0150 0.0262 0.0370 0.0487 0.0607 7.0K 0.0145 0.0272 0.0382 0.0505 0.0627 7.5K 0.0160 0.0302 0.0672 0.0557 0.0692 8.0K 0.0165 0.0312 0.0747 0.0575 0.0685 8.5K 0.0180 0.0345 0.0472 0.0607 0.0777 9.0K 0.0197 0.0355 0.0485 0.0645 0.0802 9.5K 0.0200 0.0385 0.0527 0.0677 0.0862 10.0K 0.0207 0.0395 0.0540 0.0712 0.0887 10.5K 0.0232 0.0427 0.0587 0.0767 0.0922 11.0K 0.0225 0.0437 0.0597 0.0785 0.0985 11.5K 0.0240 0.0467 0.0645 0.0835 0.1037 12.0K 0.0247 0.0477 0.0657 0.0852 0.1060 12.5K 0.0262 0.0510 0.0702 0.0982 0.1472 13.0K 0.0267 0.0520 0.0715 0.0920 0.1283 13.5K 0.0282 0.0550 0.0747 0.0977 0.3422 14.0K 0.0287 0.0577 0.0772 0.0995 0.3672 14.5K 0.0315 0.0592 0.0805 0.1053 0.1297 15.0K 0.0307 0.0602 0.0827 0.1070 0.1320 15.5K 0.0325 0.0632 0.0872 0.1105 0.1380 16.0K 0.0337 0.0642 0.0890 0.1145 0.1402 iPSC Global Broadcast Times in Seconds (NX 3.1) -- Doug Pase -- ...ucbvax!tektronix!ogcvax!pase or pase@Oregon-Grad.csnet