fileserver performance [small benchmark results]

fileserver performance [small benchmark results]

[Steve]

Some time ago, I asked:

> Has anyone benchmarked a Plan 9 file server vs. a u9fs file server?
> If so, how do they compare?

I only had 2 machines, and I couldn't decide whether to sacrifice a cpu
server or file server for my setup (I knew I wanted a terminal).

I didn't get any replies, so I thought I'd do some small benchmarks to
see for myself. I did 2 tests, for the first one I copied /lib/image/*
to /tmp, for the second, I cat'ed all the images to /dev/null.
For all tests, /tmp was not ramfs, it was ~/tmp (same disk as /lib/image).

For the u9fs tests, the file server was an SGI INDY running IRIX 5.2,
serving a Seagate ST3610N to a MIPS Magnum 3000 plan 9 terminal.

For the plan 9 file server tests, the file server was a MIPS Magnum 3000,
with a Seagate ST11200N disk (same Magnum 3000 terminal).

For each test, I repeated it 4 times, and used SGI's NetVisualizer to see
how busy the ethernet was during the test (LAN was otherwise idle).

bench u9fs:
  cd /lib/image; time cp * /tmp
	0.00u 3.70s 19.04r 	 cp dna.3c mandrill.3c swan.3 table.3c ...
	0.02u 3.70s 20.52r 	 cp dna.3c mandrill.3c swan.3 table.3c ...
	0.04u 4.30s 20.20r 	 cp dna.3c mandrill.3c swan.3 table.3c ...
	0.00u 4.64s 19.78r 	 cp dna.3c mandrill.3c swan.3 table.3c ...
	(using 30% of ethernet bandwidth)

  cd /lib/image; time cat * > /dev/null
	0.00u 0.42s 4.54r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	0.00u 0.32s 4.76r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	0.02u 0.30s 4.56r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	0.02u 0.32s 4.54r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	(using 80% of ethernet bandwidth)

bench plan 9 file server:
  cd /lib/image; time cp * /tmp
	0.02u 3.12s  8.00r       cp dna.3c mandrill.3c swan.3 table.3c ...
	0.02u 3.14s  7.88r       cp dna.3c mandrill.3c swan.3 table.3c ...
	0.00u 3.18s  7.78r       cp dna.3c mandrill.3c swan.3 table.3c ...
	0.02u 3.10s  8.14r       cp dna.3c mandrill.3c swan.3 table.3c ...
	(using 45% of ethernet bandwidth)

  cd /lib/image; time cat * > /dev/null
	0.02u 0.56s 4.04r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	0.00u 0.72s 4.06r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	0.00u 0.70s 4.10r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	0.00u 0.68s 4.02r 	 cat dna.3c mandrill.3c swan.3 table.3c ...
	(using 40% of ethernet bandwidth)

So, it seems that for the first test (reading & writing to the same disk
at the same time) u9fs took 2.5 times as long as a real Plan 9 file server.
Since the ethernet utilization was not 2.5 times greater, that should be
an indication of the efficiency of il vs. tcp.

For the second test, the times were closer together, but I'm not sure why.
Maybe the main problem with u9fs is when writing, so in the read-only test it
does much better. Again, we see that u9fs/tcp gobbles up the network (80%).

I was surprised that I didn't notice any caching effects. The times were
the same for a newly rebooted system as for re-running the benchmarks
over and over again right away.

Overall, it seems like the performance penalty of using u9fs isn't
that bad unless you do a lot of concurrent reading & writing.

Any other opinions or explanations?

	Steve

ps. Should I be using 'disc' instead of 'disk', as in the Plan 9 papers?