From: nils toedtmann <user-mode-linux-user@ni...> - 2004-11-08 16:57:26
I have a really annoying issue with uploading huge files (like iso9660s) to
a UMLed vsftpd. The problem does not occur on downloads from the UMLed vsftpd.
If this is a well-known-issue please point me to the messages in the archive,
i did not find it. Here it comes:
Download start with 10 MByte/s (it's a 100Mbit/s link) and decrease to minimum
7 MByte/s, that's perfect. Upload starts with a 10 MByte/s, too, then suddenly
1) client sends with 10MByte/s for 1 sec
2) server stops ACKing. client has to wait for server for ~ 10 secs
3) server ACKs. Goto (1)
Resulting overall rate is << 1MByte/s :-(
Server is a 2.6.9-bb1 UML (on a 126.96.36.199-skas5 host) with vsftpd. Server
load goes >10. It's still responsive (with lag of cource) via ssh. System
is mostly "100% waiting" with high "system" usage (all consumed by vsftpd)
once in a while. When i switch the inner fs from ext3 to ext2 (outer fs
is ext3 in either case) the crack down happens later - sometimes after
1GB uploaded data - but it does still happen.
At the beginning of the upload, outer "vmstat 1" shows ~10000 bo IO (0%
waiting) per second, inner "vmstat 1" alternating between ~12000 and ~8000
bo IO (0% wa). 10 seconds before crackdown, both report suddenly ~0 bo IO
(still 0% wa). When problem starts, outer vmstat shows constantly ~4000 bo
IO (~30-40% wa), inner shows ~0 bo and every once in a while a bursts of
very different size (100% wa!).
After upload, inner vmstat drops slowly to 0 bo IO, _then_ CPU drops from
100% wa to 100% idle. Outer vmstat still shows 4000bo IO (~30-40% wa) for
~20 seconds and drops then to ~0 bo IO (100% id).
No notably swapping observed all the time (host & guest), other UMLs and
host running as usual (i could compile a kernel on the host at the same
moment in reasonable time ;-)
Is that a journalling issue? Would switching fstype help?
there is no sig.