[Search for users]
[Overall Top Noters]
[List of all Conferences]
[Download this site]
| Title: | FDDI - The Next Generation |
|
| Moderator: | NETCAD::STEFANI |
|
| Created: | Thu Apr 27 1989 |
| Last Modified: | Thu Jun 05 1997 |
| Last Successful Update: | Fri Jun 06 1997 |
| Number of topics: | 2259 |
| Total number of notes: | 8590 |
1015.0. "Latency, a customer question" by EURHEP::BROWN () Tue Jul 06 1993 13:41
I just received a series of questions from a customer in Germany who is
looking to purchase a farm of AXP workstations connected together using
FDDI/GIGAswitch.
One of the questions that was asked was:
"What is the time of a memory - memory copy over FDDI
(send request -> receive data)?"
Obviously, this is dependent on the amount of data you are transfering,
the amount of time you wait for a token, etc. However, it does raise
an interesting question which is relevant for workstation farms.
What is the maximum packet per second rate that an AXP OSF/1 system can
source or sink with TCP/IP packets and 1 byte payloads? Assume that the
test setup is 2 AXP workstations, DEFTA adaptors, point-to-point connect
so one can use full duplex protocols.
I read through the Digital Technical Journal article on high performance
TCP/IP and UDP/IP, and saw very few numbers. There were a series of graphs,
one of which (page 56) I can use to estimate a pps rate for small packets of
4,500 or a latency of 225 microseconds. The packet rate seems a bit on the
low side to me, but it is hard to read a graph which is that small.
The maximum packet-per-second rate should be an indication of 1/latency.
Overall transmission latency would add wires and GIGAswitch, etc. Latency
is very important for workstation farms where long syncronization times
can kill your scalability.
Can anyone give any data as to single packet latency and 1-byte
packet-per-second rates?
Also, how different would the setup be for maximizing pps rates verus bit rates?
Thanks in advance,
Michael Brown
International Account Consultant
European High Energy Physics Group
Geneva, Switzerland
| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 1015.1 | TCP may make it hard to measure just one byte | JEDI::CAUDILL | Kelly - NaC Tech Support - 264-3320 | Tue Jul 06 1993 13:49 | 10 |
| I am no expert on this stuff, but this one item in your question got my
attention:
> What is the maximum packet per second rate that an AXP OSF/1 system can
> source or sink with TCP/IP packets and 1 byte payloads?
My experience is that TCP will take many of your single byte send()
calls and pack them up into one datalink message. So a test of this
may not be so easy to do. UDP, on the other hand, would send them out
when you think it is sending them.
|
| 1015.2 | See also BERGEN::Workstation_Farms | STAR::PARRIS | SLED is DEAD, long live RAID | Tue Jul 06 1993 15:57 | 0 |
| 1015.3 | | KONING::KONING | Paul Koning, A-13683 | Tue Jul 06 1993 17:56 | 14 |
| Incidentally, please keep in mind that latency and throughput are INDEPENDENT
parameters. It is not valid to take a latency number (e.g., 225 microseconds)
and deduce from that the throughput for that packet size.
There is of course some connection between the two. The connection may be
a very direct one IF your protocol is a request-response protocol (such as
Netware). Conversely, the connection may be very loose if the protocol is
a bulk stream protocol (e.g., DAP in file transfer mode).
An analogy: suppose you know it takes an airplane 6 hours to cross the
Atlantic, and the plane holds 300 people. What does that tell you about the
number of people that can cross the Atlantic per day? Answer: nothing...
paul
|