c_rstine@HNS.COM (Robert Stine) (03/11/91)
I'm evaluating the network performance of a single-board computer that runs ethernet and TCP/IP; the board has an 82596CA LAN controller. In one of the tests, I use the program "spray", executed from a Sun SPARC IPC, to see how well the board can handle high volume IP inputs. (I have used the default settings for ICMP ECHO requests, in which Spray hammers the target board with 1162 ICMP echo requests in frames of 86 bytes each, sent as fast as the Sun can generate them. Spray can generate either RPC or ICMP traffic. Because the board under evaluation does not run a spray daemon, I am banging it with the ICMP traffic.) The Sun and the target board are on the same MTU. For the tests, I put the MTU in loopback mode, to avoid nasty looks from other LAN users :-) What are reasonable drop rates? In "System Performance Tuning," Mike Loukides suggests that the RPC drop rate should be under 5%. Is this also reasonable for ICMP? I am observing drop rates considerably over 5% (e.g., over 20%). Thanks, Bob Stine c_rstine@hns.com
BILLW@MATHOM.CISCO.COM (William "Chops" Westfield) (03/12/91)
What are reasonable drop rates? In "System Performance Tuning," Mike
Loukides suggests that the RPC drop rate should be under 5%. Is this
also reasonable for ICMP? I am observing drop rates considerably over
5% (e.g., over 20%).
Well, I am aware of certain high performance networking devices that
consider ICMP, and especially responding to pings, a low priority task.
Can't you spray the box with the sort of packets you would like it to
handle during its eventual normal operation? RPC, UDP discard and TCP
discard all seem like better ideas...
BillW
-------mahesh@pyrhard2.pyramid.com (Mahesh Jethanandani) (03/21/91)
In article <9103111409.AA10491@hns.com> c_rstine@HNS.COM (Robert Stine) writes: > >I'm evaluating the network performance of a single-board computer that >runs ethernet and TCP/IP; the board has an 82596CA LAN controller. In >one of the tests, I use the program "spray", executed from a Sun SPARC >IPC, to see how well the board can handle high volume IP inputs. I have done similiar testing between our systems Ethernet board ( I am not sure what chip it uses) and MIPS 120 box. Here are the results. Actually these are some of the worst results, I have seen. Generally I would say anything less than 5% drop in case of RPC and anything less than 10% for spray with ICMP echo request is good enough. With spray the most important factor is the buffer you have on the Ethernet board to receive packets. Testing with RPC ... Sending 10000 packets 2457 packets (24.570%) dropped by orville Testing with ICMP echo request ... Sending 10000 packets 711 packets (7.110%) dropped by orville Testing with RPC ... Sending 20000 packets 4707 packets (23.535%) dropped by orville Testing with ICMP echo request ... Sending 20000 packets 14800 packets (74.000%) dropped by orville Testing with RPC ... Sending 30000 packets 17076 packets (56.920%) dropped by orville Testing with ICMP echo request ... Sending 30000 packets 22802 packets (76.007%) dropped by orville Testing with RPC ... Sending 40000 packets 21448 packets (53.620%) dropped by orville Testing with ICMP echo request ... Sending 40000 packets 44 packets (0.110%) dropped by orville mahesh