goldsten@cs.ubc.ca (Norman Goldstein) (01/22/91)
On page 55 of "Transputer Technical Notes", INMOS Limited, 1989, it is stated that using a crossbar connection "... introduces a 1.6 to 2 bit time delay on the signal." Is this per byte? Are there results of testing to compare hardwired connections with C004-implemented connections? Any info and pointers to info are appreciated. Thanks. -- Norman Goldstein University of British Columbia
koontz@CAPSRV.JHUAPL.EDU ("Kenneth W. Koontz, JHU/APL ", koontz@capsrv.jhuapl.edu) (01/22/91)
Norman Goldstein writes: >On page 55 of "Transputer Technical Notes", INMOS Limited, 1989, it is >stated that using a crossbar connection "... introduces a 1.6 to 2 bit >time delay on the signal." Is this per byte? Are there results of >testing to compare hardwired connections with C004-implemented connections? Way back in '89 (or so), I read the same stuff and wondered the same thing. At that point in time, Inmos stated that there was a bit delay but that transputers that supported overlapped acks (re T800s) would not experience any significant delay. Horse hockey! If you try a simple unidirectional throughput test between two T800 with a C004 and without a C004 (hardwired connections), you find the following difference in throughput rate (links at 20Mbps, using fairly large buffers >= 1Kbyte): without C004 : 1.75 Mbytes/sec with C004 : 1.33 Mbytes/sec Inmos does acknowledge this fact now, but back then it was a surprise to some even at Inmos. Things only get worse if you go through 2 C004s. I haven't verified it myself but Inmos claims it goes down to around 0.8 Mbytes/sec. Conclusion: if you have a really high-throughput problem that requires the maximum amount of throughput, don't use the C004s. P.S. I hope things like this don't occur again with the H1 and C104. Ken Koontz The Johns Hopkins University Applied Physics Laboratory Laurel, MD email: koontz@capsrv.jhuapl.edu
homeis@cs3 (01/24/91)
Norman Goldstein (goldsten@cs.ubc.ca) writes: >On page 55 of "Transputer Technical Notes", INMOS Limited, 1989, it is >stated that using a crossbar connection "... introduces a 1.6 to 2 bit >time delay on the signal." Is this per byte? Are there results of >testing to compare hardwired connections with C004-implemented connections? >Any info and pointers to info are appreciated. Thanks. The delay I mesured was 175 nanoseconds (at 10 Mbit/s). This delay has effect on every bit, that means every bit you put into a c004 has a propagation delay of 1.75 bit periods befor it comes out of the c004 again. The c004 only stores a single bit per channel, not a byte. So every sigal (bit or byte) is delayed 1.75 bit times. I think, the delay is caused by the signal regeneration, done by 5 times oversamling if the link input signal (see inmos technical note 18, connecting inmos lins, page 8). If you use T800s or other transputers with the fast link protocol (overlapping acknowledge), the delay of the C004 may slow down the data transmission because the acknowledge comes too late so that the sender can't send continously. I cannot measure this because I have only T414s and T212s. This point was discussed in early 1989. I stored these postings, if you need these old postings, please e-mail me directly. -------------- Dieter Homeister, Universitaet Stuttgart, Institut fuer parallele und verteilte Hoechstleistungsrechner (IPVR) W7000 Stuttgart 1, Azenbergstr. 12, Tel 0711-121-1342, Germany e-mail homeister@informatik.uni-stuttgart.dbp.de
homeis@cs3 (01/26/91)
Some people asked me for the old postings, so I send them to the news instead of mailing them idividually. I deleted only irrelevant lines to keep it short. ------- From: braner@batcomputer.tn.cornell.edu (Moshe Braner) Subject: Re: C004 Dynamic Switching Posted: Tue Feb 21 22:50:07 1989 ... One disadvantage of using a 2-layer C004 switching network is that, going through 3 C004's in a row, the data transfer rate on the link is reduced by a half (from about 1.8 MBytes/sec to 0.9). At least that's the case on our Niche board (T800-20s). I have it on good authority that using only 1 or 2 C004s on one link does not have this affect: the acknowledge bits get back just in time. Did anybody try that? Is there any solution for the layered approach (e.g. faster switch-chips)? ... ------------ From: davidb@inmos.co.uk (David Boreham) Subject: C004 does slow down links Posted: Thu Mar 2 20:23:23 1989 ... braner@batcomputer...... writes that he has it on good authority that one and two C004s in a T800--T800 link do not slow down the data-rate. Unfortunately, this is not correct. One C004 slows the link down by about 20%, two devices by about 40%. ... This posting is not to be regarded as INMOS customer support. From: HALLAM@vax1.physics.oxford.ac.uk (P.Hallam-Baker) Subject: Inmos links, the C004 and everything. Date: 8 Mar 89 17:31:03 GMT ... In respones to the recent letter from D.Bareham it is interseting to note that he does not explain why the C004 slows down links. The T.puter data sheet states that the T8 is capable of sending the acknowlege about 5 bit periods after the start of receiving the message. This would suggest that the C004 which introduces a 1.5 bit delay should not slow the link. Unfortunately the link is not able to begin transmission of the next byte until 5 bit cycles after receiving the acknowlege. Appologies to Dave Jefferson of CERN who worked out all this stuff for pre-empting but people seem to need to know ! ... ------- From: davidj@cernvax.UUCP (davidj) (Dave Jeffery) Subject: Re: Inmos links, the C004 and everything. Date: 13 Mar 89 07:38:23 GMT ... We are preparing a paper on Inmos Links which will be presented at the 'Computing in HEP' conference at Oxford in April. However to summarise quickly.... The T8 does not send an ack until 5-6 bit periods after receiving the start bit. And does not send the following datya byte until 5-6 bit periods after receiving an ack. Thus the cable connecting two t8's has to be very short unless you want to drop the data rate. The 5-6 figure arises because of jitter on the signal and clock sync at each end of the link. There are additinal factors due to cable delayt, buffer delay, c004 delay and which direction the wind was blowing when you made the measurements. We have derived a formula which accurately predicts the transfer rate between: t8-t8, t8-t4, t4-t8 and t4-t4 - see the proceedingd of the conference for more details.... The c004 slows down the data transmission as it resync'c the data before forwarding it, the delay is approx 1.9 bit periods at 20 Mbps link speed -------------- Dieter Homeister, Universitaet Stuttgart, Institut fuer parallele und verteilte Hoechstleistungsrechner (IPVR) W7000 Stuttgart 1, Azenbergstr. 12, Tel 0711-121-1342, Germany e-mail homeister@informatik.uni-stuttgart.dbp.de