[erlang-questions] trouble with erlang or erlang is a ghetto
Thu Jul 28 12:51:13 CEST 2011
On 28 Jul 2011, at 02:39, Richard O'Keefe wrote:
> On 28/07/2011, at 12:06 AM, Lukas Larsson wrote:
>>  http://kth.diva-portal.org/smash/record.jsf?pid=diva2:392243
> It looks as though the claim that Erlang runs out of steam at 16
> is a furphy, and that there are good ideas about how to make it
> even better.
Indeed (although I had to look up the word "furphy").
The picture on page 64, illustrating near-perfect scalability of the Big Bang benchmark on a simulated 128-core SPARC  is nice - partly because not everyone has access to the fairly pricey Simics simulations, but also because it illustrates how much of the scalability problem on multicore is about negotiating the divide between processor speed and memory bandwidth .
One may also ponder, based on this, how important it is to understand the memory access characteristics - and by extension, the dependency patterns - of the application you want to scale. These are exciting times… :)
 I learned from Simics experts that SPARC is easy to scale this way because it is a very symmetrical architecture. In contrast, many other vendors have specific architectures for their 4-, 8-, 16-core systems, and so on, so just tweaking the number of cores is not very meaningful.
 For those who haven't read the paper, the simulation assumed zero cost for memory access.
Ulf Wiger, CTO, Erlang Solutions, Ltd.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the erlang-questions