spray - 44% packet loss

[frostm]

I've got a user complaining that when he does a spray (basically a ping on steroids) to any other host in his vlan that he gets almost 50% packet loss. But when he connects those hosts to a $40 hub, he gets no packet loss at all.

Background information:

1. It doesn't matter which host he does the spray from or to. It happens regardless of whether the two hosts are on the same switch or not.

2. sh proc - less than 2% cpu

3. sh system - peak traffic 5%

4. When I do a ping -t I get no losses. In fact, from looking at the switch, I can't find evidence of loss of frames or excessive errors. Only the return from the spray command at the host, reveals the losses.

So why can't my 5500's keep up with a $40 linksys hub??

[jwitherell]

I'd suspect a duplex mismatch / autonegotiation problem. Try fixing/nailing his port to half duplex. Since it works OK on a hub, that indicates the NIC may want to stay at half duplex.

Tell us more about this "spray". It sounds like an interesting utility. How can you do it, or is there shareware out there for it?

[frostm]

OK I gathered a bit more information.

Found this from the man pages on spray

"spray is not useful as a networking benchmark as it uses unreliable connectionless transports, (upd for example). spray can report a large number of packets dropped when the drops were caused by spray sending packets faster than they can be buffered locally (before the packets get to the network medium)."

When I read that, I wanted to go tell the user to get lost. But he's got these servers the size of refigerators, and they supposedly cost more than I'll make in the next two decades. So I need to have my ducks in a row first.

So I had a buddy do a spray on another vlan. He said he got no packet loss spraying from one slow machine to another. But he got 58% packet loss spraying from a fast machine to another fast machine.

I'm having the user check his nic properties, and hopefully get rid of autnegotiation on his nic. Then I'll do the same on the switch. Then I was planning to clear my counters on the switch and set up a tcpdump into a file on both the sending a recieving hosts. Hopefully that might give us some clues when we do the next set of sprays and pings.

Any thoughts on how I would best set up my logging/debugging on the switch to gather more info?