> On Jan 18, 2006, at 7:09 PM, Samuel RICHARD wrote:
>
>> I was wondering what happens with LAM MPI tcp rpi if the TCP send
>> buffer
>> (inside the kernel) is full (the application is computing faster
>> than the
>> network sends packets).
>> I think the socket API must return an error when the application is
>> trying
>> to write data to the socket but how does LAM behaves in such a case?
>> Does LAM tries to re-send data after a defined delay?
>
> Check the man page for write(2).
>
> In some cases, LAM will put the socket in a blocking mode, and if
> more data is written than the kernel will handle, write() will simply
> block until all the data is able to be sent.
>
> In other cases, LAM will put the socket in a non-blocking mode. If
> more data is written than the kernel can handle, write() will return
> the number of bytes actually written, and LAM keeps track of what
> bytes of the message still need to be written.
>
> In short -- LAM handles the return values from write() properly to
> ensure that the entire message is able to be sent.
>
> Hope that helps.
>
That helps a little but what I want to know is in case of non-blocking
mode : when does LAM write bytes of the message that have not been
written?
Does it retries regurlarly? and in this case what is the re-try delay?
|
