[erlang-questions] erlang vs. ibm power8 in the cloud: super-high single-system parallelism

[Felix Gallo]

'runabove.com' (said to be a subsidiary of OVH?) recently announced an IBM
Power8 cloud server offering in which you can spend about $1 an hour to get
176 threads on and 48G.  Signing up for that lab thing appears to give you
about a day's worth of runtime credit at that rate.  So in the interests of
erlang-questions-list citizen journalism, I signed up for the test.

Each individual thread is a little less speedy than those of us in the x86
world may be used to, but after horsing around with installing tar (!?),
ncurses, and a development environment for a bit, and then installing a
recent erlang via kerl, these servers are gigantic, fast, and...

[root@REDACTED admin]# erl
Erlang/OTP 17 [erts-6.1] [source] [64-bit] [smp:176:176] [async-threads:10]
[kernel-poll:false]

Eshell V6.1  (abort with ^G)
1>

erlang recognizes all 176 threads.  Nice!

sysbench for this box:

[root@REDACTED admin]# sysbench --test=cpu --cpu-max-prime=100000 run
--num-threads=176
sysbench 0.4.12:  multi-threaded system evaluation benchmark
[...]
Test execution summary:
    total time:                          2.8208s
    total number of events:              10000
    total time taken by event execution: 443.7155
    per-request statistics:
         min:                                  7.52ms
         avg:                                 44.37ms
         max:                                118.15ms
         approx.  95 percentile:              48.13ms

For comparison, a generic cloud x86 8-core box:

root@REDACTED:~# sysbench --test=cpu --cpu-max-prime=100000 run
--num-threads=8
sysbench 0.4.12:  multi-threaded system evaluation benchmark
[...]
Test execution summary:
    total time:                          38.7198s
    total number of events:              10000
    total time taken by event execution: 309.5309
    per-request statistics:
         min:                                 29.89ms
         avg:                                 30.95ms
         max:                                 79.24ms
         approx.  95 percentile:              34.12ms

No idea how production-ready these boxes are, and don't have a
single-system highly parallel erlang throughput benchmark handy to drop in
there to see how they do, but if you want a glimpse of the multicore
future, these could be pretty fun to play around with.  One could imagine
video streams and databases being pretty insane.

F.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20141021/491a3c0a/attachment.htm>