ORNL tests 2

Tests at ORNL 10/30-11/3/95

WARNING: These results are from preliminary tests.

Configuration

One GigaNet,Inc ATM board installed in rogue.css.ornl.gov, a 4-node (MP3/NIC B), 128 MB/node, (can run either NIC A or NIC B kernel and one service node Paragon. OSF 1.3.0. New libatm.a (10/23/95) Second ATM board in Intel's XPE (mpdot1.ccs.ornl.gov), 16-node GP/NIC A, same software.

Sandia has provided a FORE ATM switch with OC12 and OC3 interfaces, a BBN long link emulator, and an AT&T EMMI .

Compute node atm message-passing tests

Tom Dunigan ran a number of message passing tests using a compute node on each Paragon communicating through ATM. We booted rogue with a NIC A kernel to better emulate the SC 95 configuration. We ran echo tests, sink tests, and exchange tests to characterize message transfer time and bandwidth as a function of message size. (Click here for description of tests and see the test results from 10/19.)

Click here for compute-node plots.

Slightly slower throughput and higher jitter was observed when using a service node on each Paragon. On 11/3/95, we did see some full duplex performance, two concurrent and opposing sink tests achieved an aggregate date rate of 82 MB/s.

Here is a summary of compute-node communications, including using two Paragons (NIC A).

time(us) Bandwidth (MB/s) Nic A NX 36 90 atm-echo 140 20 atm-sink 56 58 Nic A (2 Paragons) atm-echo 140 20 atm-sink 56 55 Nic B NX 25 171 atm-echo 125 30 atm-sink 41 68

PVM/ATM tests

Phil Papadopoulus has worked on porting, testing, and tuning PVM for the ATM API.

10/30/95. PVM over ATM between two Paragons yielded a throughput of 436 KB/s and a latency of 14 ms.
Running a PVM test between the two Paragons over Ethernet (UDP/IP) yielded a throughput of 306 KB/s and a latency of 56 ms!. By comparison, PVM between two Sun workstations over Ethernet (UDP/IP) yielded a throughput of 568 KB/s and a latency of 3.5 ms.
10/31/95. PVM thru OC12 switch yielded 2 ms and 2 MB/s using 16KB blocks.
11/3/95. Bill Shelton ran material science grand challenge code over PVM between the two Paragons using both Ethernet (257.8 s) and ATM (249.1). A second test case (62 energies) took 1576.3 seconds on Ethernet and 1339.2 seconds over ATM.

ATM OC3 delay tests

Tom Pratt and Alden Jackson have designed a series of tests (see test configurations ). Here are results from 10/31/95 tests using two NIC A Paragons and the echo test. Time is 1-way time to transmit a 4-byte message, and bandwidth is measured with a 60KB message. The delay is only in one direction. Tests from OC12 into OC3 would hang for messages larger than 8 KB. We believe these failures are caused by the FORE switch (version 3.4.2).

Connection time(us) Bandwidth (MB/s) direct 136 21 oc12 switch 139 21 oc12-oc3 147 12 (at 8KB) oc12-3+1us 160 12 oc12-3+1ms 660 7 oc12-3+10ms 5169 1 oc12-3+20ms 10171 1 oc12-3+50ms 25157 0.6 oc12-3+100ms 50161 0.2 PVM (8 KB frags) oc12 switch 2560 2 oc12-oc3+1ms 2860 0.8 oc12-3+50ms 27000 0.1 oc12-3+100ms 54000 0.07 oc12-3+1ms 8 KB PVM + errs 2^-22 4940 0.7 2^-27 2850 0.9

IP/ATM and ATM video tests

The test configuration for IP over ATM and ATM video is depicted here.

10/30/95, Prohaska tests new kernel on mpdot1 and initial IP/ATM.
10/31/95, ATM/IP: ifconfig/netstat working on mpdot1
11/1/95, TCP and UDP working over ATM between rogue and mpdot1. ttcp and UDP latency tests worked
11/1/95, Initial ATM video tests (AT&T Emmi (Q=25) and Paragon)