On Wed, Jan 13, 2010 at 12:24 AM, Jason Gunthorpe
jguntho...@obsidianresearch.com wrote:
I'd actually think that the decline caused by exhausting credit on
the initiator side is far worse than ib_post_recv.. If ib_post_recv
really is noticable then maybe just force loading the RQ after XX
Hello,
I've been redirected to this list from OpenFabrics' web site, so let me apologise
right away if this is not the right place for this question.
We are writing an application that requires high throughput and low latency. More
specifically, we need to take advantage of features like
On 1/13/2010 12:06 PM, Werther Pirani wrote:
Hello,
I've been redirected to this list from OpenFabrics' web site, so let
me apologise right away if this is not the right place for this question.
We are writing an application that requires high throughput and low
latency. More
These patches allow to have a specific QoS for Lustre traffic.
- Patch 1 adds a new Lustre port space to the rdma_cm kernel module
- Patch 2 defines a new Lustre service ID (matching the new RDMA CMA port
space) and a new 'lustre' keyword in the OpenSM QoS policy parser.
- Patch
This patch is only given as a reference.
Index: b/lnet/klnds/o2iblnd/o2iblnd.c
===
--- a/lnet/klnds/o2iblnd/o2iblnd.c
+++ b/lnet/klnds/o2iblnd/o2iblnd.c
@@ -2503,7 +2503,7 @@ kiblnd_startup (lnet_ni_t *ni)
This patch adds a new service ID for the Lustre ULP as well as the
lexer and parser bits for the policy definition simplified syntax.
The QoS policy parser now supports the following 'qos-ulps' keywords:
lustre : 0 #default SL for Lustre
lustre,
Sometimes nodes are discovered by the opensm with their default
name before real host name was resolves and updated in the HCA.
Since Mellanox ConnectX HCA do not support NodeDescription
change trap (as part of trap 144) we need an option to refresh
NodeDescription from console.
Signed-off-by:
This patch adds a new port space for use by Lustre traffic.
This, along with patches to OpenSM and Lustre, allow to define a
specific QoS for lustre.
In general, I don't like the idea of application port spaces for QoS
support. Can't this be done using port numbers in the existing port
The impact of the ib_post_recv() call is small but measurable. I'll
publish measurement results soon.
Interesting. If there is overhead you could amortize it by keeping some
number of spare buffers posting, and batching your calls to
ib_post_recv() (eg post a list of 8 buffers every 8
In general, I don't like the idea of application port spaces for QoS
support. Can't this be done using port numbers in the existing port
space?
I agree. If setting QoS requires a kernel patch for each application, I
think we've messed up QoS somehow. There needs to be a way to control
On 17:34 Wed 13 Jan , Eli Dorfman (Voltaire) wrote:
Sometimes nodes are discovered by the opensm with their default
name before real host name was resolves and updated in the HCA.
Since Mellanox ConnectX HCA do not support NodeDescription
change trap (as part of trap 144) we need an
On Wed, Jan 13, 2010 at 08:23:27AM +0100, Bart Van Assche wrote:
On Wed, Jan 13, 2010 at 12:24 AM, Jason Gunthorpe
jguntho...@obsidianresearch.com wrote:
[ ... ]
Also, I couldn't tell for sure from a cursory examination of the
patch, but the initiator must be designed to not stall
On Wed, Jan 13, 2010 at 7:16 PM, Jason Gunthorpe
jguntho...@obsidianresearch.com wrote:
On Wed, Jan 13, 2010 at 08:23:27AM +0100, Bart Van Assche wrote:
On Wed, Jan 13, 2010 at 12:24 AM, Jason Gunthorpe
jguntho...@obsidianresearch.com wrote:
[ ... ]
Also, I couldn't tell for sure
Should be 132 rather than 131 according to IANA IP protocol IDs
Signed-off-by: Hal Rosenstock hal.rosenst...@gmail.com
---
diff --git a/include/rdma/rdma_cm.h b/include/rdma/rdma_cm.h
index c6b2962..1c5e589 100644
--- a/include/rdma/rdma_cm.h
+++ b/include/rdma/rdma_cm.h
@@ -67,7 +67,7 @@ enum
Should be 132 rather than 131 according to IANA IP protocol IDs
Signed-off-by: Hal Rosenstock hal.rosenst...@gmail.com
Acked-by: Sean Hefty sean.he...@intel.com
---
Thanks. I just noticed this this morning after reviewing the luster patches.
You beat me to the patch.
diff --git
On Wed, Jan 13, 2010 at 07:57:26PM +0100, Bart Van Assche wrote:
On Wed, Jan 13, 2010 at 7:16 PM, Jason Gunthorpe
jguntho...@obsidianresearch.com wrote:
On Wed, Jan 13, 2010 at 08:23:27AM +0100, Bart Van Assche wrote:
On Wed, Jan 13, 2010 at 12:24 AM, Jason Gunthorpe
Hi Sasha,
On Tue, Jan 12, 2010 at 2:14 PM, Sasha Khapyorsky sas...@voltaire.com wrote:
Hi Hal,
On 14:10 Thu 31 Dec , Hal Rosenstock wrote:
to --details option
Signed-off-by: Hal Rosenstock hal.rosenst...@gmail.com
---
diff --git a/infiniband-diags/man/ibqueryerrors.8
On Tue, Jan 12, 2010 at 5:35 AM, Sasha Khapyorsky sas...@voltaire.com wrote:
On 08:38 Thu 07 Jan , Hal Rosenstock wrote:
Change appears to be for switches to always rely on this bit rather
than only when VLCap is 1. I wonder if there are any switches with
VLCap 1 that don't set the
On Tue, Jan 12, 2010 at 5:43 AM, Sasha Khapyorsky sas...@voltaire.com wrote:
On 08:40 Thu 07 Jan , Hal Rosenstock wrote:
On Tue, Jan 5, 2010 at 6:10 AM, Sasha Khapyorsky sas...@voltaire.com wrote:
This splits QoS related port parameters setup over switch external ports
and end ports.
On Tue, Jan 12, 2010 at 5:11 AM, Sasha Khapyorsky sas...@voltaire.com wrote:
Hi Ralph,
On 13:43 Mon 11 Jan , Ralph Campbell wrote:
Its been 2 months since I posted this.
What is the current status? Last comment I saw was from Ira
saying either add this functionality to ibportstate or
Added by Ralph Campbell ralph.campb...@qlogic.com
in commit 3af8e05e01a92bb504c8e806c0c81dc5ef246f7f
Signed-off-by: Hal Rosenstock hal.rosenst...@gmail.com
---
diff --git a/infiniband-diags/man/ibportstate.8
b/infiniband-diags/man/ibportstate.8
index fea860e..c4c1b2a 100644
---
-RDMA_PS_SCTP = 0x0183
+RDMA_PS_SCTP = 0x0184
The old wrong value has been there since we first merged this. So
changing it means anything using this in userspace would be broken. But
is PS_SCTP exported to userspace? And as far as I can tell PS_SCTP is
not implemented (this
The old wrong value has been there since we first merged this. So
changing it means anything using this in userspace would be broken. But
is PS_SCTP exported to userspace? And as far as I can tell PS_SCTP is
not implemented (this line is the only reference to RDMA_PS_SCTP in the
whole kernel).
RoCEE was chosen for these patches in accordance with the terminology
used in the proposal discussed by the OFA. We have no problem
resubmitting these patches using iboe so that you can push them into
2.6.34. We will need to stay aware of any future name recommendations
by relevant
Technically, I believe it's available to userspace, though the librdmacm
doesn't
define it. I'd be surprised if anything uses it.
Wouldn't trying to use RDMA_PS_SCTP hit the
default:
return -EPROTONOSUPPORT;
case in cma_get_port()? So one couldn't actually do
Wouldn't trying to use RDMA_PS_SCTP hit the
default:
return -EPROTONOSUPPORT;
case in cma_get_port()? So one couldn't actually do anything with
RDMA_PS_SCTP? Or is there some way to avoid binding?
You're right. There isn't any way to actually do anything with it. All
Add ib_uverbs_get_mac() to be used by ibv_create_ah() to retirieve the remote
port's MAC address from the remote port's GID. Port link layer is also returned
by ibv_query_port().
Is this really to map the GID of a remote port to its MAC address, or is the
mapping for a local port?
If it's the
27 matches
Mail list logo