[rescue] Solaris vs sparclinux

[Meelis Roos]

> Hmm, it would be interesting to install Solaris on a spare disk in T2000 
> and run some bencharks.

Well, I did that - grabbed Solaris 11.2 and installed it on identical 
disk.

Debian unstable with 4.0.0 kernel and my custom configurationf for that 
machine (left out most drivers I will never need, left out some features 
I would not use).

Server is T2000, 1 GHz, 32 threads, 16G RAM, 10kRPM SAS disks. Bare 
hardware, no virtualization if any kind (no ldoms, zones, containers).

> And, should I run single-disk Solaris on ZFS or journaled UFS for this?

ZFS was the only option.

> Also, what is the latest freely available Solaris and does it let me 
> publish the benchmarks?

Nope, the license does not allow publishing of benchmarks without 
Oracles permission. So no numbers, just subjective impressions.

The results are mixed again.

First, one of the tools that I often wait after is git (for big repos 
like Linux kernel), so I first tested different git scenarios (with git 
2.3 on both systems)- initial remote clone, local clone, git fsck (rare 
but slow if it happens), checkout. This led to more filesystem tests - 
copy and rm of the tree.

In these tests, Linux was clearly superior - on the average about twice 
as fast as Solaris. The only big exception was git fsck that was 
CPU-bound and here Solaris was about twice as fast as Linux. Also, rm -r 
of a big tree was slighlty faster on Solaris. Maybe ext4 is just so much 
faster on single disk...

So I tried lmbench3 to test more than filesystes. Notable difference was 
Bcopy where niagara-optimized Solaris implementation was about 3 times 
as fast as Linux. It appears that Linux glibc has implementations for 
ultrapsarc-T* variants too but Debian has disabled them because they saw 
some bug some time ago, and it hasn't been reported or investigated 
until now, and Debian-packaged glibc uses ultrasparc-III level of 
optimizations only.

There were 2 other trends: first, context switch was slightly but 
consistently faster on Solaris (more on no dirty data(?), less so on 
more dirty data) and local communication latencies were faster on Solaris on 
the average. There were exceptions but this was the general trend. Also, 
pipe and AF_UNIX bandwidth was higher in Solaris.

Another trend was that anything related to filesystems was about twice 
as fast on Linux, same as in macrobenchmarks.

So I fired up hackbench (CPU scheduler benchmark that uses socket pairs 
for ping-pong of data between tasks). I expected Solaris to win here but 
was suriprised to see Linux clearly winning.

Another real-life-like test for shell scripts was simplistic spawning of 
shells, like

time for i in `seq 10000`; do sh -c true; done

Here again Linux was many fimes faster.

So the results are mixed but for most of my real-life-like test on 
command line, sparclinux still seems to be faster.

Now if the glibc ultrasparc-T* optimizations could get fixed... it has 
just risen into the radar of the main sparclinux developer, DaveM.

-- 
Meelis Roos (mroos at linux.ee)