[gentoo-user] Slow network transfers ... lost interrupts because of clocksource?

["Stefan G. Weichinger" <lists@xunil.at>]

I am back from my visit at a customer where I installed a new and shiny
gentoo server for running VMs (KVM).

Currently I don't have access as my VPN only works from my static IP at
home (my router seems to be offline right now ... and I am still away
from office for the weekend) so I can't check details now ...

basically:

I tried to copy/rsync some file with ~8GB over a gigabit connection ...
from old to new server. Checked ethtool for gigabit, looked ok. I always
saw the behavior that the transfer started rather fast and slowed down
within minutes. Let's say ~50 MB/s in the start and then down to maybe 2
or so. That is way from the expected throughput with such new hardware.

The NICs in the new server are BCM-something, Broadcom, using the tg3
Tigon module (exact model not available right now as mentioned above).

In "dmesg" I see lines like

hpet1: lost 1 rtc interrupts

which makes me wonder if that leads to the lousy performance (btw, I
fear slow virtualization performance as well).

Dealing with HPET I checked for kernel support and also added kernel
options to GRUB:

hpet=force clocksource=hpet

which lead to

# cat /sys/devices/system/clocksource/clocksource0/current_clocksource
hpet

I am unsure if I should further investigate things around this HPET-issue?

My thinkpad here shows "tsc" as clocksource ... good/better ?

What direction to go? force or disable HPET?

-

btw I saw the same slowdown on 2 NICs: one connects the server to the
LAN via GB-switch, the other connects the 2 servers directly via
crossconnect-cable (dedicated for backups).

I tried rsync with various options, scp, and even some tar/netcat-combo ...

The system runs gentoo sources 3.10.7, amd64 ...

-

Thanks for any thoughts on this!

Stefan