RE: [coLinux-devel] Benchmarks: coLinux 0.6.3 vs. coLinux 0.7.1-hn12vs. FC3 2.6.9-1.667

[Marc E. F. <mef@CS.Princeton.EDU>]

Hello Henry,

> coLinux runs on a single machine and have only up to 90% of CPU time.
> Standard benchmarks runs not long enough to check the real speed.  The
> clock inside coLinux is synched with windows clock.  The bad thing for
> your test is, that coLinux ask the host for accurate time, every you
> call gettimeofday().  Yes, it goes from Linux into host OS and back, to
> give you the time.  In this case you lose some milliseconds.

How does it go into the host for gettimeofday() calls?  Is the communication
asynchronous (i.e., using a separate thread running in Windows to provide
the answer), or is it primarily reading a shared page with the data
synchronously that is updated every now and then by the Windows OS?

> That's why you can only compair some long working stuff, for sample
> kernel builds and copy disk with dd.

Good feedback.

> Remember that, if you use a real partion for cobd
> (\Device\HarddiskX\PartitionY), then coLinux currently does a sync disk
> access (waits for IO ready).  That is 5x or more slower as a root
> filesystem in a file.

For our usage scenarios it doesn't make sense to use a real partition.  We
are using a file backed filesystem.

> CoLinux 0.6.x and 0.7.x runs on different kernels.  If you compair it
> agains native Linux, you should run the same kernel version.  I think
> kernel 2.6.12 has some internal latency overheads.  Linux kernel
> devolopers made some experiments on limit per process time.  That can be
> slower for some task switch.  But I'm not known more about this.

Yes, we need to compare to 2.6.12 vs. 2.6.9.  However, given the
aforementioned timing problems with gettimeofday(), I don't think that is
our primary problem.

> For your networking speed: CoLinux has no special load balancing between
> Linux and Windows, all bytes goes through a pipe of 4096 bytes passage
> page.  If that pipe is still overloaded (with your clock-request) then
> it has no more time for byte transfer.  Every block that should transfer
> from or to coLinux will wakeup a sleeping colinux-daemon on that pipe.
> That is the condition, why somebody incrase the process priority for
> networking daemons to speed up networking.

Is there a way to dynamically change this priority from within coLinux?

> CoLinux is optimiszed for speed to running on the Linux side without
> blocking the host system.  With some changes, speed would incrase for
> transfers (network, block devices, console).  The current design use a
> pipe concept for communication to the host.  In this case the single CPU
> must run on both end of the pipe in same time.

Interesting.

Marc