mark@sickkids.UUCP (Mark Bartelt) (12/23/88)
Before describing the problem, here's the configuration: VAX/750 running
4.3bsd, with a Kinetics FastPath 2 connecting the ethernet to an AppleTalk
network. We have both NCSA telnet (version 2.1) and the PacerLink software
(version 5.0a) available for people to use for connecting Macs to the VAX.
Compared to telnet, the PacerLink software seems a bit sluggish.
That in itself isn't totally surprising, since the Pacer software has a lot
more stuff buried inside it, for doing file transfers and such. Just out
of curiosity, I used /etc/ping to see how long it would take Macs running
telnet and PacerLink to respond. Using 500-byte ping packets, the times
for the Pacer software were more than three times longer than those for
NCSA telnet. Again, not altogether surprising. One thing that seemed a
little peculiar was the fact that, on a 200-packet test, not only were the
average times larger, but so were the standard deviations (as percentages
of the means). The times reported by "ping" were 93ms +- 17% for telnet,
whereas the Pacer times were 300ms +- 44%.
However, the *real* oddity is what we saw when we actually looked at the
ping times. For telnet, the values look pretty much as one would expect,
with values more or less randomly scattered around the mean:
89 112 158 121 105 76 100 89 82 75
76 80 81 111 114 110 81 78 119 91
114 75 101 79 91 95 122 100 91 78
82 93 81 116 82 79 112 80 80 121
130 83 80 90 123 124 96 90 95 91
93 114 120 81 86 92 80 80 80 81
77 87 87 112 163 82 76 117 91 115
104 76 81 85 112 119 93 85 81 83
83 98 98 76 93 87 82 76 96 79
89 77 83 109 81 83 121 109 80 95
But with the Pacer software running in the Mac, there's an unexpected
periodicity, with near-monotonicity within each period. Take a look at
these numbers (read down the columns for sequential ping times):
231 334 236 148 362 284 340 204 445 364
270 390 295 185 382 340 399 225 466 384
290 389 344 205 420 394 464 246 90 407
311 412 402 227 442 447 522 266 155 423
336 433 452 248 477 518 546 304 211 448
369 470 511 268 482 525 112 325 246 485
107 490 551 289 520 101 90 346 299 502
154 515 90 310 103 161 124 366 370 91
209 100 104 331 156 231 147 404 424 142
281 165 128 341 210 291 167 425 344 195
Does anyone have a clue (or would anyone like to hazard a guess) as to
what's going on here?
Mark Bartelt UUCP: {utzoo,decvax}!sickkids!mark
Hospital for Sick Children, Toronto BITNET: mark@sickkids.utoronto
416/598-6442 INTERNET: mark@sickkids.toronto.edusbm@PURDUE.EDU (12/24/88)
The question was about round-trip times reported by ping: > But with the Pacer software running in the Mac, there's an unexpected > periodicity, with near-monotonicity within each period. The reason for this behavior is pretty simple. It looks like the Pacer software uses the default socket listener, which posts an event when a packet arrives, and checks for events about every half second, but not exactly every half second, so that some packets will arrive just before the event queue is checked, yielding the actual packet turnaround time of 90-103ms, and others will arrive just after the queue was checked, yielding the longest delay (caused by waiting for the software to check the queue again) of 466-551ms. Ping sends packets every second, so you see a nice progression as the phase between ping and the Pacer software slowly changes. NCSA telnet, apparently, either checks for events very often (perhaps once every 30ms, judging from the variation of the round-trip times), or uses its own socket listener to respond immediately, interrupt-style, when packets arrive. Steve Munson sbm@Purdue.EDU sbm@Purdue.CSNET ----------