[Swift-devel] nearly, but not quite optimal performance on ranger from ranger under coasters

[Glen Hocky]

In the rare case where more coasters become active than you expected, 
shouldn't you fill them up from the submitted queue? or at least have 
that as an option?

Mihael Hategan wrote:
> On Wed, 2009-04-29 at 16:44 -0500, Glen Hocky wrote:
>   
>> Hi Everyone,
>> Today I tried to run jobs of length ~45 minutes. I ran on Ranger from 
>> Ranger with
>>     <profile namespace="karajan" key="initialScore">50</profile>
>>     <profile namespace="karajan" key="jobThrottle">5</profile>
>>
>> This submitted 51 jobs to the queue, of which 38 went to the run state, 
>> meaning there should be 608 available coasters. I  had
>>     
>>> Progress:  Selecting site:2569  Active:431  Finished successfully:9
>>>       
>> For ~30 minutes until some jobs started finishing. Then I had a stead 
>> decline of activity, see below.
>> I assume this is happening because the extra 177 coasters that were 
>> available timed out durning this 45 minutes and new ones did not become 
>> available, but that could be wrong.
>>
>>
>> How should I tune my settings to fix this?
>>     
>
> There is no tuning to fix this. In the current scheme, if
> coastersPerNode > 1, there will always be more workers requested than
> jobs submitted. If sufficient jobs are not subsequently submitted by
> swift (and that requires the first round to finish), some workers will
> die having done nothing.
>
>