You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@river.apache.org by Patricia Shanahan <pa...@acm.org> on 2011/08/01 01:23:36 UTC
Re: Future plans
On 7/31/2011 1:43 AM, Peter Firmstone wrote:
...
> * TaskManager - improve concurrency and remove the dependency on
> Task.runAfter() in River code.
...
I'm playing with implementing a subclass of ThreadPoolExecutor that
modifies the number of threads based on the mean number of tasks over
some recent time period. That way, the number of threads will not be a
long term bottleneck, but threads will not be created and destroyed for
short-lived bursts.
I have an on-going concern, for all performance and concurrency work in
River, that we have neither benchmarks nor access to real installations
to instrument and observe.
Patricia
Re: Future plans
Posted by Tom Hobbs <tv...@googlemail.com>.
To be honest, I've got no interest in Internet-visible services or
discovery. Not that I want to discourage other people working on
them. If that's your itch, then scratch away! :-)
Greg's new wiki page said one of the things that particularly hits
home with me; "If it were easier to get up and running..."
Absolutely, that's one of the things that I want to get sorted. I've
spoken about it before and usually been told that it's the job of
downstream projects to make 'stuff' easier. I kind of resent that
attitude. There is no denying that there is a lot of truely brilliant
stuff in downstream projects, that's no reason to leave using vanilla
River as such a headache though.
Other things that come to mind are the Gradle build and the
distributed transaction manager.
We should probably start thinking about a new roadmap as well. Yes,
this is another "vision" thing, but now we're a TLP I think we're
expected to have one. ;-)
Cheers,
Tom
On Mon, Aug 1, 2011 at 1:10 AM, Peter Firmstone <ji...@zeus.net.au> wrote:
> Patricia Shanahan wrote:
>>
>> On 7/31/2011 1:43 AM, Peter Firmstone wrote:
>> ...
>>>
>>> * TaskManager - improve concurrency and remove the dependency on
>>> Task.runAfter() in River code.
>>
>> ...
>>
>> I'm playing with implementing a subclass of ThreadPoolExecutor that
>> modifies the number of threads based on the mean number of tasks over some
>> recent time period. That way, the number of threads will not be a long term
>> bottleneck, but threads will not be created and destroyed for short-lived
>> bursts.
>>
>> I have an on-going concern, for all performance and concurrency work in
>> River, that we have neither benchmarks nor access to real installations to
>> instrument and observe.
>
> I'm having the same issue, I've noticed some Sun T2000's becoming available
> on ebay for around $1000, 32GB ram, 8 cores which appear as 32 cpu's in the
> system.
>
> The only problem might be Oracle's licensing, but this is the cheapest most
> relevant hardware I can find at present.
>
> Object based benchmarks should be easy enough to create using MultitreadedTC
> (included in test/lib), but this doesn't apply to integration testing and I
> suspect the current qa suite might have some concurrency issues too.
>
> Peter.
>
>
Re: Future plans
Posted by Peter Firmstone <ji...@zeus.net.au>.
Patricia Shanahan wrote:
> On 7/31/2011 1:43 AM, Peter Firmstone wrote:
> ...
>> * TaskManager - improve concurrency and remove the dependency on
>> Task.runAfter() in River code.
> ...
>
> I'm playing with implementing a subclass of ThreadPoolExecutor that
> modifies the number of threads based on the mean number of tasks over
> some recent time period. That way, the number of threads will not be a
> long term bottleneck, but threads will not be created and destroyed
> for short-lived bursts.
>
> I have an on-going concern, for all performance and concurrency work
> in River, that we have neither benchmarks nor access to real
> installations to instrument and observe.
I'm having the same issue, I've noticed some Sun T2000's becoming
available on ebay for around $1000, 32GB ram, 8 cores which appear as 32
cpu's in the system.
The only problem might be Oracle's licensing, but this is the cheapest
most relevant hardware I can find at present.
Object based benchmarks should be easy enough to create using
MultitreadedTC (included in test/lib), but this doesn't apply to
integration testing and I suspect the current qa suite might have some
concurrency issues too.
Peter.