[Pvfs2-developers] Re: the halloween bug fixed

[Scott Atchley]

On Oct 9, 2007, at 9:31 AM, Sam Lang wrote:

>> Why not have tcp simply not register on each connection, but on  
>> each new client. The handle_new_connection() function is only used  
>> on the server.
>
> I'm not sure that helps solve the problem.  Even with only a few  
> hundred clients, with MPI-IO, a new process on each node is going  
> to connect to the server.  Over time, a different client port  
> number could be used, so the count of "peers" could still grow into  
> the hundreds of thousands.

I see what you mean. I was not thinking of the large number of  
available ports that the client could use and that would generate  
separate client peernames. :-)

Scott