Subject: Re: Some Xen performance measurements
To: Martti Kuparinen <martti.kuparinen@iki.fi>
From: Manuel Bouyer <bouyer@antioche.lip6.fr>
List: port-xen
Date: 04/01/2005 18:53:50
On Fri, Apr 01, 2005 at 12:29:43PM +0300, Martti Kuparinen wrote:
> Hi!
> 
> I tested NetBSD/xen domain0 and domainU (3.99.2) to get some idea
> if this whole thing is suitable to run several low-CPU-intensive
> hosts (Linux and xBSD) on a single 1U Dell PowerEdge 750 with 4 GB
> of RAM.
> 
> My ultimate plan is to use bridge(4) and vlan(4) to build complex
> networks using a single hardware (less noise, heat and electricity).
> I'll write a howto similar to my RAID-1 for root filesystem document
> once everything is up and running.
> 
> I haven't received those Dells yet so I tested the same setup (but less
> domUs running bacause I only have 1 GB of RAM) on my desktop Dell
> Dimension 8400.
> 
> 
> NetBSD 3.99.2 (XEN0) #3: Fri Apr  1 10:46:57 EEST 2005
>         root@localhost:/usr/src/sys/arch/i386/compile/XEN0
> total memory = 122 MB
> avail memory = 119 MB
> mainbus0 (root)
> cpu0 at mainbus0: (uniprocessor)
> cpu0: Intel (686-class), 2793.13 MHz, id 0xf34
> ...

Is it a hyperthreaded P4 ?
This can make some difference; if HT is enabled in BIOS xen will switch to
SMP mode, spreading domains on different CPUs. I didn't run any benchmarks
to see if this gives some benefits or not.

> was this:
> 
> rm -rf /usr/src/sys/arch/i386/compile/GENERIC
> cd /usr/src/sys/arch/i386/conf
> config GENERIC
> cd ../compile/GENERIC
> time make dependall
> 
> Results
> =======
> 
> 1 x NetBSD 1024 MB GENERIC
> --------------------------
> 446.60 real       405.61 user        31.97 sys
> 
> 1 x NetBSD 128 MB XEN0
> ----------------------
> 490.51 real       411.75 user        63.14 sys
> 
> 1 x NetBSD 128 MB XENU (vnd with 2 GB file-as-disk)
> ---------------------------------------------------
> 492.76 real       415.54 user        61.42 sys
> 
> 1 x NetBSD 128 MB XENU (direct access to the wd0x partition)
> ------------------------------------------------------------
> 493.13 real       415.02 user        61.98 sys
> 
> 4 x NetBSD 128 MB XENU (vnd with 2 GB file-as-disk)
> ---------------------------------------------------
> 1633.38 real      1396.94 user       177.58 sys
> 1641.86 real      1404.19 user       195.83 sys
> 1642.72 real      1403.38 user       196.25 sys
> 1633.69 real      1394.42 user       176.68 sys
> 
> 4 x NetBSD 128 MB XENU (direct access to the wd0x partitions)
> -------------------------------------------------------------
> 1626.80 real      1396.27 user       174.35 sys
> 1642.40 real      1405.37 user       196.97 sys
> 1627.71 real      1395.80 user       174.98 sys
> 1642.59 real      1406.99 user       195.47 sys
> 
> Remarks and questions
> =====================
> 
> - One would assume that running 4 VMs would take 4x the compilation time
>   compared to 1 VM. Any ideas why the time is only 1600 instead of 2000?

Probably for the same reason make -j4 is faster than make: with only one gcc
at a time, the CPU is idle when waiting for the disk. With make -j4 or
4 domU, there should always be something runnable for the CPU.
Would be interesting to know how fast a make -j4 is.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--