[libfabric-users] connection-less send/recv with verbs
arun.ilango at intel.com
Mon Jul 17 10:30:49 PDT 2017
The server has to be started before the client and it waits for the first message from client.
The handshake lets the server and the client know each other's fabric addresses which is then used in the data transfer calls.
> On my system, I see a bunch of “sockets;ofi-rxm” providers (no verbs;ofi-rxm or similar).
This shouldn't happen. Can you run the following command and share the logs?
FI_LOG_LEVEL=debug FI_LOG_PROV="verbs,ofi_rxm" fi_info -p verbs
From: Maurizio Drocco [mailto:drocco at di.unito.it]
Sent: Saturday, July 15, 2017 2:33 AM
To: Ilango, Arun <arun.ilango at intel.com>
Cc: libfabric-users at lists.openfabrics.org
Subject: Re: [libfabric-users] connection-less send/recv with verbs
Thank you Arun,
> server: fi_rdm_pingpong -p verbs -s
> client: fi_rdm_pingpong -p verbs <server_verbs_interface_ip_address>
This partially resolves the issue, but it still does not work if the server starts with some delay with respect to the client (that is the case in my code).
The following gets stuck:
fi_rdm_pingpong -p verbs 192.168.2.253 & sleep 1; fi_rdm_pingpong -p verbs -s 192.168.2.253
Maybe a correlated point: my code does not perform the initial “handshake” for filling AVs, it simply fills them “statically”.
Is there any specific reason for which the handshake (as is performed by any sample code I saw so far) should be preferable?
> You can also try using the ofi_rxm provider which provides a connectionless interface to verbs. This is a layered provider which runs over FI_EP_MSG interface of verbs provider.
On my system, I see a bunch of “sockets;ofi-rxm” providers (no verbs;ofi-rxm or similar).
One of them is as follows:
Should I try this one?
University of Torino, department of Computer Science Via Pessinetto 12, 10149 Torino - Italy
More information about the Libfabric-users