[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Xen-ia64-devel] Scp big file to xen0 will fail
- To: <xen-ia64-devel@xxxxxxxxxxxxxxxxxxx>
- From: "Zhang, Xiantao" <xiantao.zhang@xxxxxxxxx>
- Date: Thu, 6 Apr 2006 13:07:24 +0800
- Delivery-date: Wed, 05 Apr 2006 22:09:29 -0700
- List-id: Discussion of the ia64 port of Xen <xen-ia64-devel.lists.xensource.com>
- Thread-index: AcZZNa+fZB7i4RQ7RSy4OsuryCa4ngAAi7Qg
- Thread-topic: Scp big file to xen0 will fail
Currently, scp a big size file (IG or bigger) to xen0 will fail in
midway without VTi or xenU exist. The phenomena is as below.
1.Boot machine to xen0
2.start some applications such as xend, vncserver and so on.
3.scp a big size file from remote machine to xen0, and will see scp
status to be stalled and xen0 is slower and slower. At last, the whole
My analysis about it:
When it happens, we can see the most of cpu time was occupied by pdflush
in xen0. But pdflush thread was only scheduled when system has small
amount of free buffer memory and it will flush buffer pages to disk.
This operation will do flush_tlb_all, and as to ptc.e of flush_tlb_all
needs xen to emulate. The emulation of xen has very low performance. It
will flush all vhpt entries and all tlbs of LP. Then, the sshd will get
less chance to be scheduled. From client view, network broken and xen0
hangs. So the root cause should be low efficiency of ptc_e emulation .
In order to prove this thought, I have done several experiments on it.
1. Extend xen0's memory to 640M from 512M, this will reduce times of
pdflush operation heavily. So the ptc.e emulation also decreases
2. Disable vhpt of xen0, if so, xen don't need to do flush vhpt in ptc.e
The either of two methods can resolve the issue. But they are just for
workaround not the final solution. Anyway, in order to fix the bug, we
should find better solution to improve performance of ptc.e emulation .
Please give comments :)
My machine status:
Platform : tiger4
Please give comments.
Xen-ia64-devel mailing list