Re: Benchmarks of Linux based stubdom

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Jiageng Yu, le Fri 18 Nov 2011 00:17:15 +0800, a écrit :
>    Since the periodic achievement is obtained in Linux based stubdom project, I
> am very interesting in the benchmarks of Linux stubdom. To have a good
> comparison with mini-os based stubdom, I want to do the same benchmarks as you
> did in http://www.xen.org/files/xensummitboston08/SamThibault_XenSummit.pdf.
> Please offer me the tools and methods you had used to measure the mini-os based
> stubdom. Any details would be thankful.

There is not too much fancy in there :)

- Inb (Kcy) is maybe the fancy part: the measurement was made with
  assembly bits, basically: rdtsc; inb; rdtsc, i.e. something like:

  unsigned long t1, t2;
  __asm__ volatile("rdtsc" : "=A" (t1))
  inb(0x80);
  __asm__ volatile("rdtsc" : "=A" (t2))

- Boot time is from xm create up to "foo login:" prompt.
- Disk performance was measured as seen from the guest, I don't remember
  exactly how, probably with a simple program working on /dev/xvdb
  with O_DIRECT. The CPU% was read from xm top.
- Net performance was measure as seen from the guest, probably with
  netperf. The CPU% was also read from xm top. Note the use of the e1000
  virtual device, which iirc performed best at the time.

Samuel