[EMAIL PROTECTED] wrote:
> On Tuesday 08 February 2005 12:11, Nuutti Kotivuori wrote:
>> I am wondering what is the status of ubd=mmap these days

[...]

>> The same questions about /dev/anon, with and without ubd=mmap.
>
> Both things are not working... ubd-mmap, also, cannot work, so it's
> going to be deleted from the code. No matter which version you
> use...

Okay. I tried ubd=mmap and got a kernel panic, so I kind of figured.

>> The reason I am wondering is more because ubd=mmap might hide
>> another problem we are seeing,
>
> Please provide a description... no idea about this.

We have a problem with abysmal disk-IO performance with UML in certain
cases. There have been many different problems overlaying each other,
and we still haven't got the real cause sorted out.

There are two different ways the IO performance problem can manifest
itself. The primary debugging tool has been 'dstat' running on the
host to show IO statistics. When simply dd'ing from /dev/zero to a
file inside UML and everything goes well, the numbers seen on the host
side with dstat correlate with the throughput seen inside UML, and
there is basically only write activity happening in the magnitude of 8
to 12 MB/s (depending on the drive tested).

In one problem case, the dstat on the host shows a constant read and
write load, with the read being perhaps a fourth of the amount
written. The write load and the total throughput inside UML is stable
somewhere between 300 to 600 kB/s, depending on the drive tested. This
would look like it is doing partial writes all the time, requiring it
to read as well as write.

In the other problem case, the dstat on the host shows only write load
of good magnitude, around 16 MB/s. But it is writing a whole lot more
than what happens on the guest side. In fact, in one case, 15 times
more (7.5 GB vs. 500 MB). The throughput on the guest side is a little
better, 1.5 MB/s to 3.0 MB/s depending on the drive, but the drive is
screaming like mad when this is happening. This would look like it is
rewriting the same thing over and over again.

However, the amount of variables is still too great to determine where
the problem indeed is. For files, there have been problems with
mismatching block sizes in host and guest filesystems, double ext3
filesystems with data=ordered, cow files versus images flattened with
uml_moo (which makes sparse files) versus manually unsparsified files,
LVM2 (device-mapper) partitions (both raw and files inside LVM
partitions), guest kernel 2.4 versus 2.6, etc. etc.

One thing has been constant though - adding a loopback device between
the file/partition to be accessed and pointing uml to the loopback
device has solved all our problems. It makes the IO on the host side
seem bursty (periods of silence followed by 50MB/s bursts), but
performance is decent and no problems detected. I expect it to break
all write ordering / sync expectations however, but those need to be
investigated better anyway.

Anyway, still digging more information to pinpoint the problem.

-- Naked



-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://ads.osdn.com/?ad_id=6595&alloc_id=14396&op=click
_______________________________________________
User-mode-linux-user mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-user

Reply via email to