Avery Ching commented on GIRAPH-128:

Thanks for taking a look.  I forgot to upload the original (rb only for that 
one), hence part 2. 

The main motivation for the obscure case is that it would make debugging 
simpler.  We often see errors like serverX:portY, and can use portY to figure 
out which mapper to look at.  For example, currently the default starts at 
30000.  If I see an error from 30001, then I know to go to mapper 1 to see it's 
problem.  And so on and so forth.  If I am running a 900 mapper job then if 
it's 31001 or 32001 then I still know to look at mapper partition 1.  If 
instead I had a 100 as the constant, then if it's 30101, I have to check both 
mapper 1 and mapper 101.  With up to 20 retries per port, we can handle at 
least 20 simultaneous jobs running on a single machine that have the same 
mapper partition id.  First of, that is probably unlikely.  But even if it does 
happen, 20 is probably more than an one machine would handle.  By the way, port 
retries are very fast (so I wouldn't worry to much about collisions).

Let me resubmit without the whitespace changes and making MAX_BIND_ATTEMPTS 
> RPC port from BasicRPCCommunications should be only a starting port, and 
> retried
> --------------------------------------------------------------------------------
>                 Key: GIRAPH-128
>                 URL: https://issues.apache.org/jira/browse/GIRAPH-128
>             Project: Giraph
>          Issue Type: Improvement
>    Affects Versions: 0.1.0
>            Reporter: Avery Ching
>            Assignee: Avery Ching
>         Attachments: GIRAPH-128.2.patch
> Currently Giraph uses a basic port + the task partition to get the RPC port.  
> This doesn't work well for when there are multiple Giraph jobs running 
> simultaneously in the same Hadoop cluster (port conflict).  At the same time, 
> it is nice to use this simple algorithm because it makes it very easy to 
> debug problems (you can find the troublesome mapper from the RPC port name).  
> I will be proposing a simple scheme to retry with another port.  I will round 
> the total number of mappers up to the nearest power of 10 (let's that that 
> number Z).  Then I will increment the port number by Z, retrying up to 20 
> tries.  If you have enough ports, this scheme would guarantee that up to 20 
> mappers / node would be supported.  It should be sufficient for most 
> clusters.  At the same time, we still maintain the easy debugging method 
> since you it's still easy to figure out the mapper partition from the port 
> (port % Z = map partition). 

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: 
For more information on JIRA, see: http://www.atlassian.com/software/jira


Reply via email to