[coLinux-devel] Performance Tuning

[Digital I. Inc. <ok...@di...>]

Performance Tuning : rough idea

Okay, coLinux is running on many PCs. A lot of problems, but running. Then it is
time to start a discussion about how to tweak coLinux.
I think coLinux has problems mainly in three criteria.

  - X
  - File access
  - Memory shortage


Before starting, it is wise that we discuss many possibilities from many aspects.
Your idea, comments, anything is welcome.
Of course, other than them, we have many problems and many points to improve.
Your opinion is definitely welcome.

                                 ---- Okajima.


-------------

  - X
 I think it is not so slow even at this moment. But you can improve it by using
virtual inter process communication (IPC) between coLinux and Windows through
a passage, insted of using virtual TCP link currently used.
Some people say that making X run on coLinux natively, it speeds up. but I am
doubtful about this. because if do so, you can not use Windows 2D graphics 
accelerator driver.

  - File Access
 This is slow. and I think direct sector access to NTFS partition though NT
sector read driver would help.

  - Memory shortage
 coLinux has only very little memory. There is a lot of ways to solve this,
and your idea is welcome. I think creating virtual UMA(using main DRAM as VRAM - 
used mainly with on-board VGA.) device on NT is one way. It's memory is not
going to be swapped, probably.
Another idea is, using this. http://linuxcompressed.sourceforge.net/
Second, coLinux uses memory doubly. for example, it caches sector data doubly.
I mean, NT caches them. and coLinux caches same data.
Someday coLinux can run multiple instance, and the problem will become more
serious. because, there are many same ELF, same .so and same something....
 between VMs.
XEN ( http://www.cl.cam.ac.uk/Research/SRG/netos/xen/ ) is working on this problem.
In next version ( XEN 1.3 ), they have a plan to use shared ELF and .so between
VMs. And even it shares chache buffer.
And Jeff Dike, the author of UML, has same plan.