[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Xen-devel] OOM killer problem

Thank you for reply.

>> In my examination, vbd thread became often OO-states.

> We'll need more info. You mention the vbd thread, so it is it usually
domain0 that ends up in OOM state? How much memory 
> do you a allocate each domain0 and each other VM?

My envrionment is
 X86 (Pentium 4), 2.5GB Mem and IDE and SCSI disk.

I allocate 1GB memory for domain0 and 256MB memory for each domainU (4VMs).
Sweps are allocated 2.5GB on IDE for domain0, and No swaps for domain U
the dd program run on each domainUs,
VM disks for domainU are used pysical partition on SCSI (ex. /dev/sda1
/dev/sda2 /dev/sda3/ /dev/sda4).

Oom-killer wakes up for each domainU, and kills each deamon such as sshd.
Thereafter, it kills login and bash, so return login console.

Above OO-states shows state in vbd-back driver.
I seem that each variable counts following state.
  oo_req is blockd requests.
  rd_req is pended read requests.
  wr_req is rended write requests.

In the case of three VM,  OO state is not so occured.
However, In the case of four VM, OO state is occured enormously.

So, I seem that I/O is interfered by I/O from other VM, and so slower.
Therefore, I/O request is pileing up, and so OOM killer is waked up.

I create observation program for VBD I/O requests.
This patch will through after few days.

>> So, I think that this cause is that I/O request is not processed 
>> faster.
>> This problem is found in the case file systems is slower by some Web 
>> page.

>Can you clarify what you mean here?

In past, OOM killer is became the problems all over the file system. 
  ex ext3, nfs.
I research these problem not enough, but I think to occur sisutation
 which is similler to explain above.

Thank you,
Satoshi UCHIDA

Attachment: smime.p7s
Description: S/MIME cryptographic signature

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.