Procurve 4160GL Poor Performance to 100Mbs Client

[Charles Hooper]

I just upgraded from an HP Procurve 4000M core switch to an HP 4160GL 60 port 10/100/1000 core switch (switch is 2 weeks old). Gigabit device to gigabit device performance is good to excellent (depending on the computer hardware), but a pull from a gigabit device (running either Win NT 4.0 Server or Windows 2000 Server) to an 100Mb device takes a very long time, at a rate of 10-15Mb/s. Transferring a 129MB file (Windows 2000 SP 4) takes roughly 90 seconds, compared to 15-19 seconds when using only the 4000M core switch at 100Mbs.

The switch indicates that all devices are operating at full duplex and maximum speed of the network cards, and there are no errors reported at the switches. Oddly, if I pull the same file from a gigabit capable Red Hat Enterprise 3.0 server running Samba 3.0 to my 100Mb/s client, the copy completes in 18 seconds. Only the Windows servers/computers are affected by gigabit to 100Mb speed difference, while the Linux server has no problem on similar hardware.

If I connect another switch between the 4160GL and the 100Mb client through a gigabit crossover cable (fiber or copper), the performance problem disappears. The second switch can be either the HP 4000M, a Dell 3248, or a Dell 3348 (or probably anything else).

Only pulls are affected - pushes from one server/computer to another operate normally. Ethereal running on one client computer (CH) shows that there were dropped packets and DUP ACKs when pulling the test file from the gigabit capable server.

See the attached Word document for test results.