Version 1.2.c
Installation Guide
- July 2007
+ August 2007
==============================================================================
Table of contents
Open Fabrics Enterprise Distribution (OFED)
- Version 1.2.c-10
+ Version 1.2.c-11
Release Notes
- July 2007
+ August 2007
===============================================================================
o Documentation
Notes:
-1. mlx4 driver is in beta quality.
+1. mlx4 driver is in GA quality.
2. cxgb3 driver is in technology preview state.
3. The Virtual NIC (VNIC) driver is presented as a technology preview on OFED 1.2.
4. All other OFED components are of production quality.
- InfiniHost III Ex (MemFree: fw-25218 Rev 5.2.000
with memory: fw-25208 Rev 4.8.200)
- InfiniHost III Lx (fw-25204 Rev 1.2.000)
- - ConnectX IB (fw-25408 Rev 2.1.000)
+ - ConnectX IB (fw-25408 Rev 2.2.000)
For official firmware versions please see:
http://www.mellanox.com/support/firmware_table.php
- core (the verbs providers)
===============================================================================
-4. Fixed Bugs since OFED 1.2.c-9
+4. Fixed Bugs since OFED 1.2.c-10
===============================================================================
-- Kernel 2.6.22 is now supported
-- Add libsysfs and libstdc++ 32-bit dependency for openmpi on PPC64.
-- iser: kfifo.c now taken from upstream kernel.
-- cma: Added module parameter to allow adjusting the response timeout for
- situations when the default timeout is not long enough
+- mlx4 drivers were not created on kernel 2.6.22
+IPoIB:
+- Leak in ipoib_transport_dev_init
+- Add interrupt moderation support for ipoib
+- Fixed kernel oops in IPoIB download
+
Bugs fixed in mlx4 (ConnectX driver) are reported in mlx4_release_notes.txt.
Open Fabrics Enterprise Distribution (OFED)
- IPoIB in OFED 1.2 Release Notes
+ IPoIB in OFED 1.2.c Release Notes
- June 2007
+ August 2007
===============================================================================
4. DHCP Support of IPoIB
5. High Availability (HA) Service
6. The ib-bonding driver
+7. Bug fixes and Enhancements since OFED 1.2
===============================================================================
1. Overview
* The bondX name must not contain characters which are disallowed for bash
variable names such as '.' and '-'
+
+===============================================================================
+7. Bug fixes and Enhancements since OFED 1.2
+===============================================================================
+- Add interrupt moderation support for ipoib
+- NAPI is avilable using a module parameter
+- Fixed a leak in ipoib_transport_dev_init
+- Fixed kernel oops in IPoIB download
+
Open Fabrics Enterprise Distribution (OFED)
mlx4 in OFED 1.2.c-10 Release Notes
- July 2007
+ August 2007
===============================================================================
mlx4_core 0000:08:00.0: is not supported by the installed driver.
mlx4_core 0000:08:00.0: The driver supports Command i/f revision 2.
-This release was tested with FW 2.1.000
+This release was tested with FW 2.2.000
===============================================================================
3. Tested ULPs
o MPI (MVAPICH and Open MPI)
o IPoIB UD and CM
o SDP
- o SRP was partially tested
+ o SRP
+ o RDS was partialy tested
===============================================================================
4. Fixed Bugs and Enhancements
===============================================================================
- o FMRs are now supported
- In addition:
- - MTT entries now written by driver directly (WRITE_MTT firmware command
- no longer used) -- avoids cache-line write race conditions due to FMRs
- - MPT table now allocated from coherent memory, to avoid race conditions
- between FMRs and firmware.
- o query_qp now returns the RQ capabilities to userspace.
- o Prevent possible data corruption when multiple gather segments are used.
- This fix requires to write s/g segments and the inline segment (for MSI/GSI)
- in reverse order into the WQE.
- o Eliminate double-kfree in mlx4_mr_alloc error flow.
+o Added support for interrupt moderation for create cq to support ipoib.
+o Display the following device information via sysfs: board_id, fw_ver,
+ hw_rev and hca_type. As a result ibstat presenting all entries.
+o Fixed destroy special QPs.
+o Fixed the work completion opcode for RDMA_READ operation
+
===============================================================================
5. Known Issues
o APM
o SQD
-- ibstat does not present all entries. Use ibv_devinfo instead.
-
- To load the driver on machines with 64KB default page size UAR bar must be
enlarged. 64KB page size is the default of PPC with RHEL5 and Itanium with
64KB page size enabled.
Open Fabrics Enterprise Distribution (OFED)
- OSU MPI MVAPICH-0.9.9, in OFED 1.2 Release Notes
+ OSU MPI MVAPICH-0.9.9, in OFED 1.2.c Release Notes
- June 2007
+ August 2007
===============================================================================
2. Software Dependencies
3. New Features
4. Bug Fixes
-5. Known Issues
-6. Main Verification Flows
+5. Bug fixes since OFED 1.2
+6. Known Issues
+7. Main Verification Flows
===============================================================================
- Error report fixes
- PKEY support fix
+===============================================================================
+5. Bug fixes since OFED 1.2
+===============================================================================
+New version mvapich-0.9.9-1451
+
+- mpirun_rsh fixes:
+ * Check that we do not overflow command line size
+ * better handling for accept
+- Fix for memory allocations whose size is bigger than 2^31-1
+- Fix for multiple communicator problem
+- Fix for MPI_Finalize segmentation fault
===============================================================================
-5. Known Issues
+6. Known Issues
===============================================================================
- A process running MPI cannot fork after MPI_Init. Using fork might cause a
segmentation fault.
===============================================================================
-5. Main Verification Flows
+7. Main Verification Flows
===============================================================================
In order to verify the correctness of OSU MPI, the following tests and
parameters were run.
ib_ipath=n
ib_sdp=y
ib_srp=y
-ib_iser=n
+ib_iser=y
ib_ehca=n
vnic=n
-rds=n
+rds=y
ib_bonding=n
librdmacm_devel=y
librdmacm_utils=y
rds_tools=n
-dapl=n
-dapl_devel=n
+dapl=y
+dapl_devel=y
ibutils=y
mvapich=y
Version: OpenFabrics Enterprise Distribution (OFED) 1.2
Repo: git://git.openfabrics.org/~ofed_1_2/management.git (release)
git://git.openfabrics.org/~halr/management.git (development)
-Date: June 2007
+Date: August 2007
1 Overview
----------
Open Fabrics Enterprise Distribution (OFED)
- RDS in OFED 1.2 Release Notes
+ RDS in OFED 1.2.c Release Notes
- June 2007
+ August 2007
===============================================================================
2. Supported Platforms
3. Installation & Configuration
4. New Features
+5. Bug fixes and Enhancements since OFED 1.2
===============================================================================
1. Overview
===============================================================================
4. New Features
===============================================================================
-RDS socket send buffers are now per socket, not per flow, and the new recv
+RDS socket send buffers are now per socket, not per flow, and the new receive
buffer value is per socket. So we should use the core stack sndbuf and rcvbuf
values instead of offering a separate interface.
for f in {r,w}mem_{default,max}; do
echo $nr > $sys/$f
done
+
+===============================================================================
+5. Bug fixes and Enhancements since OFED 1.2
+===============================================================================
+1. rds_poll() always returns POLLOUT. If you try to send and get
+ EWOULDBLOCK because the remote receiver is congested then you don't get
+ to wait for POLLOUT to be raised before sending again -- it's always
+ raised. This is worked around by implementing some kind of exponential
+ back-off while retrying the send. This can be done by increasing the
+ timeout given to poll(). At each poll() timeout expiry the send is
+ tried again.
+
+ This idea of waking poll() waiters on congestion notification hopes to
+ cut down on the latency between when the send could succeed and when the
+ poll() timeout hits and the send is retried. If we receive a congestion
+ bitmap update from the remote node we wake poll() waiters, giving them a
+ chance to retry their send.
+
+2. Set max send and receive scatter/gather list size to 2 in the QP attributes.
+ The second sge used for RDS header.
+
+3. Added rs_poll_flag to rds_socket structure to indicate the back pressure
+ status and send space availability.
+
+ rds_poll set POLLIN in case of released back pressure, send space became
+ available and if there are messages in the receive queue.
+ set POLLOUT if send buffer is available
+
+4. Added MODULE_VERSION.
projects / compat-rdma / docs.git / commitdiff