Skip to contentSkip to conversation

Notes on XRC

DISCUSSION Shlomo Pongratz 2 participants 5 messages Mar 25, 2013 – Mar 26, 2013 4 messages not in the archive InfiniBand

lore
Share & tools
lore
#0 lore raw

Hi Sean,

I'm experimenting XRC with a real application and I want to share my thoughts.

A natural way to pass the SRQ's QPN is to use the "private_data" of the "rdma_conn_param", however this area is limited to 56 octets and since each QPN occupies 3 octets that leaves us with 18 SRQ. I suggest to increase the size of the private data, or to find a more convenient way to pass the SRQ's QPM, one that doesn't use extra ports or out-of-band communication (socket).

Another issue is of backwards compatibility of applications. Assume an application uses a well known port, and want to allow both new XRC enabled clients and old RC clients. It seems that this is possible using the "rdma_create_ep" API with appropriate "rdma_addrinfo" one that "rdma_getaddrinfo" resolved with QP type "IBV_QPT_XRC_RECV" and one with "IBV_QPT_RC" one the same port (I hope this is allowed). A client on the other side will first try to connect to the XRC and if it fails will fall-back to RC.

However, a server usually uses "rdma_create_id" to associate an event handler with an id and then bind the id to an address, and accept connection requests asynchronously, but with these API, AFAIK there is no way to pass the required hints.

I assume I can use the "private_data" of the "rdma_conn_param" to distinguish between XRC enabled clients and servers and old ones, but it doesn't seems clean, as it is implicit and not explicit.

Best regards,

S.P.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#2 lore raw
In reply to a message not in the archive
Quoted 6 lines

A natural way to pass the SRQ's QPN is to use the "private_data" of the "rdma_conn_param", however this area is limited to 56 octets and since each QPN occupies 3 octets that leaves us with 18 SRQ. I suggest to increase the size of the private data, or to find a more convenient way to pass the SRQ's QPM, one that doesn't use extra ports or out-of-band communication (socket).

This is an issue for the IBTA to deal with and requires increasing the size of a MAD.

Quoted 8 lines

Another issue is of backwards compatibility of applications. Assume an application uses a well known port, and want to allow both new XRC enabled clients and old RC clients. It seems that this is possible using the "rdma_create_ep" API with appropriate "rdma_addrinfo" one that "rdma_getaddrinfo" resolved with QP type "IBV_QPT_XRC_RECV" and one with "IBV_QPT_RC" one the same port (I hope this is allowed). A client on the other side will first try to connect to the XRC and if it fails will fall-back to RC.

The intent behind rdma_getaddrinfo is to allow multiple QP types to be returned based on user input.

Quoted 4 lines

However, a server usually uses "rdma_create_id" to associate an event handler with an id and then bind the id to an address, and accept connection requests asynchronously, but with these API, AFAIK there is no way to pass the required hints.

Correct - the older rdma cm APIs are not as flexible. That's why new APIs ended up being added.

Quoted 3 lines

I assume I can use the "private_data" of the "rdma_conn_param" to distinguish between XRC enabled clients and servers and old ones, but it doesn't seems clean, as it is implicit and not explicit.

I don't think I'm fully following your private data idea.

- Sean
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw
In reply to a message not in the archive
Quoted 7 lines · Sean
Quoted 6 lines

A natural way to pass the SRQ's QPN is to use the "private_data" of the "rdma_conn_param", however this area is limited to 56 octets and since each QPN occupies 3 octets that leaves us with 18 SRQ. I suggest to increase the size of the private data, or to find a more convenient way to pass the SRQ's QPM, one that doesn't use extra ports or out-of-band communication (socket).

This is an issue for the IBTA to deal with and requires increasing the size of a MAD.

I see, I guess this will an obstacle.

Quoted 10 lines
Quoted 9 lines

Another issue is of backwards compatibility of applications. Assume an application uses a well known port, and want to allow both new XRC enabled clients and old RC clients. It seems that this is possible using the "rdma_create_ep" API with appropriate "rdma_addrinfo" one that "rdma_getaddrinfo" resolved with QP type "IBV_QPT_XRC_RECV" and one with "IBV_QPT_RC" one the same port (I hope this is allowed). A client on the other side will first try to connect to the XRC and if it fails will fall-back to RC.

The intent behind rdma_getaddrinfo is to allow multiple QP types to be returned based on user input.

So it is O.K. to have both XRC and RC QPs with conjunction to

Quoted 6 lines
Quoted 5 lines

However, a server usually uses "rdma_create_id" to associate an event handler with an id and then bind the id to an address, and accept connection requests asynchronously, but with these API, AFAIK there is no way to pass the required hints.

Correct - the older rdma cm APIs are not as flexible. That's why new APIs ended up being added.

Can you tell me which new API will give me the ability to use both address hints and asynchronous connection requests handling.

Quoted 5 lines
Quoted 4 lines

I assume I can use the "private_data" of the "rdma_conn_param" to distinguish between XRC enabled clients and servers and old ones, but it doesn't seems clean, as it is implicit and not explicit.

I don't think I'm fully following your private data idea.

This is just a workaround I been considering in case using "hints" and asynchronous connection request can't work together. If I want to have servers and clients that need to be backward compatible with RC servers and clients. And assuming the old ones didn't used the private data in the connection request and in the accept. Than the presence of the private data will indicated that these are XRC enabled clients and servers.

S.P

Quoted 1 line

- Sean

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#6 lore raw
In reply to a message not in the archive
Quoted 6 lines

This is just a workaround I been considering in case using "hints" and asynchronous connection request can't work together. If I want to have servers and clients that need to be backward compatible with RC servers and clients. And assuming the old ones didn't used the private data in the connection request and in the accept. Than the presence of the private data will indicated that these are XRC enabled clients and servers.

I don't see how an app can 'just switch' from RC to XRC without having coded for the possibility of XRC. XRC introduces too many new concepts. --
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#8 lore raw
In reply to a message not in the archive
Quoted 7 lines · Sean
Quoted 6 lines

This is just a workaround I been considering in case using "hints" and asynchronous connection request can't work together. If I want to have servers and clients that need to be backward compatible with RC servers and clients. And assuming the old ones didn't used the private data in the connection request and in the accept. Than the presence of the private data will indicated that these are XRC enabled clients and servers.

I don't see how an app can 'just switch' from RC to XRC without having coded for the possibility of XRC. XRC introduces too many new concepts.

The communication is just one part of the application. It is acceptable to prepare for both XRC and RC and decide ad-hoc which transport to use, there is no intention to switch back and forth between the different transport. The nature of the connection is to be established with the connect-accept handshake, and it is not going to be changed as long as the connection linger. So what I need is to accept connection requests asynchronously and decide which transport module to uses.

S.P.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html