during the testing DRBD with using Intel NIC and e1000e (Intel 82571EB Gigabit Ethernet, version: 1.0.15-NAPI) driver I have identified very strange behavior. When I started replication for a few volumes (NIC with e1000e was destination) I had problem with running commands via SSH on this machine from other host. Those commands were performed very long. Replication was slow, as well. So I have changed hardware interface (on the same motherboard) which is supported by the e1000 driver and after that everything worked correctly and I did not have problem with slow running SSH commands.
Next I have prepared some test with using Samba, iSCSI and I have received this results:
Debian Lenny, kernel 18.104.22.168 32b:
iSCSI IET (e1000e):
block i/o: 54.7 write
block i/o: 62.7 read
file i/o: 54.3 write
file i/o: 58.5 read
iSCSI IET (e1000):
block i/o: 78.7 write
block i/o: 67.1 read
file i/o: 76.3 write
file i/o: 60.5 read
Is hardware with e1000e worse then hardware with e1000? Have you noticed similar behavior?
I have tried e1000e with MSI and without MSI interrupts, a few version of this driver, with enabling and disabling protocol offload via ethtool but without results.
My main problem is very slow commands via SSH when e1000e is using via e.g. DRBD. Maybe something with interrupts?
Thank you in advance