Re: [Myrinet] bogus page

Le Mardi 29 Janvier 2002 19:37, vous avez écrit :
> Salut Thomas
>
> Thomas Guignon wrote:
> > Hello,
> > Sometime our MPI application makes the gm driver write something like
> > this: GM: LANAI[0]: port=2 bogus page SEND ptr = 0x40667020 FAILED 
> > (maxidx=32767)
>
> It happens when the Lanai is trying to resolve a virtual address and this
> address is not in the registration hash, i.e. is not registered. There is 2
> causes for that, a good one and a bad one. The bad one signals a bug in the
> application, the send buffer covers a page (at least one) that is not
> registered or allocated with gm_dma_malloc. It can happen on the receive
> side also. The ptr is the virtual page address. The second case is when the
> GM port is closed when messages are in flight: the port closure will
> deregistred automatically all registered memory, but it may happen in the
> middle of a large message send processed on the Lanai at that time. The
> fist pages are Ok, the rest of the buffer appears deregistered and you have
> this message. That would happen when the process dies, segfaults, killed by
> a signal, etc.
> It's a good timing as I have just finished a change for GM-1.5.1 that would
> remove the good case: if the port is closed, no messages are printed in the
> log. So in GM-1.5.1, this type of message would indicates a bug, 100% sure.
>
> Now, if you see this message when the application is running and the port
> is open (and stays open), I would like to run your MPI code to see what's
> happening.
>
> > I think this was caused by the fact that the
> > buffer was in swap instead of physical memory, correct me if I am wrong.
>
> Registered memory is locked in physical memory, and to register pages force
> them to swap out if they are swapped, so the buffer cannot be in swap.
>
> > Does --gm-can-register-memory option forces buffer to be in DMA-able
> > memory?
>
> Are you using mpich-gm 1.2..X ? there is no more "--gm-can-register-memory"
> in 1.2.1..X

Bonjour Patrick
We are using mpich-1.2.1..7 so sorry for the stupid question on 
--gm-can-register-memory

For our application I think that most of the cases are the first one but I'am 
not really sure.
The problem I have is that the error messages does not have time stamp to 
know when the error append: the message does not appear in /var/log/message 
but only with dmesg.


A+ 
-- 
Guignon Thomas, [email protected]
Tel: 33 1 69 18 95 93
Fax: 33 1 69 18 95 96
ARTABEL S.A.
69, rue de Paris
91400 Orsay FRANCE