[erlang-questions] cheap low-power boards

[Martin Scholl]

Hello all,

[hoping this is still interesting for the ppl at erlang-questions@,
rabbitmq-discuss@]


Ben Hood wrote:
> Martin,
> 
> On Mon, Dec 29, 2008 at 1:27 PM, Martin Scholl <ms@REDACTED> wrote:
>> Ben Hood wrote:
>> [snip]
>>> Obviously there are some standard benchmarks such as throughput and
>>> latency, but these are generally used to compare different releases
>>> against each other.
>> I think, latency is more interesting than throughput. I am not really
>> familiar with RabbitMQ and its distributed capabilities -- imho it would
>> be particularly interesting to see how RabbitMQ scales on these kind of
>> embedded stuff (comparing performance/watt with standard hardware).
> 
> Sure, but if I was really worried about latency, I don't know if would
> go down the embedded route.
[I am not an eco activist and I don't want to evangelize, and no, I am
not paid by Intel for doing this:]

Some numbers: compiling erlang on an ATOM 330@REDACTED takes about 2/3
times longer than on an Intel E2160 @ 1.8ghz -- but the atom requires
just about ~10% of the power of the E2160 ==> The Atom "outperformes"
the E2160 clearly regarding performance / watt.
In other words: it is worth considering to replace servers with
(several) ATOM-based systems as long as you deploy a software that can
handle this.

For the stuff we have tested, I can say: there are not many applications
which definitely demand a non-atom processor. I am clearly not an expert
in all things RabbitMq. I guess, RabbitMq's performance is dominated by
the network technology used and not the hosts' cpus. Therefore, RabbitMq
is a candidate for atom-based systems.

> 
>> Our results are quite promising:
>>
>>> Defining what constitutes a benchmark is to an extent a subjective
>>> task and hence will depend on what your use case is.
>> I am interested in "green computing" and especially results of
>> performance / watt ratios of real world scenarios. This is my motivation
>> behind all this.
> 
> Maybe you could look at an application that does not constantly stress
> the middleware - hence having something light weight would not cost
> much when it's not doing anything and hence been *greener*. Having
> said this, I think this may be a question of how long is a piece of
> string -  for example, you could also compare bang for for energy buck
> by running virtualized machines.
Just to be sure: I don't want to evangelize. Having played with the ATOM
processor I am quite sure, there is more room for it than just the
embedded world. That is all I like to research.

[snip]

> In the Rabbit client library there is a utility called MulticastMain,
> which, if you run it the --help flag, will tell you about all of the
> configuration options you have to test latency and throughput from a
> client perspective.
Ok, will post some numbers soon.



Martin