Linux Kernel,
The PCI-SCI Drivers for the Dolphin Scalable Coherent Interface
v1.2-1 has posted at vger.timpanogas.org:/sci/pci-sci-1.2-1. This
release corrects several hardware related bugs, and corrects
several previously reported build and performance problems related
to the RedHat 7.1 Fischer Release.
These drivers are released under the GNU public license, and are
freely downloadable and re-distributable. These drivers are provided
in both .tar.gz and RPM formats.
Please direct any questions, bug reports, or comments to either
[EMAIL PROTECTED] or [EMAIL PROTECTED]
Analysis of performance problems related to the gcc 2.96 compiler on the
fischer release indicate that there are some problems with mixing
powertools RPMs and egcs compatibility modules on a base release
that may result in poor code generation. We wiped the entire
RedHat 7.1 Fischer system, and performed a clean install with just
the base compiler, and none of the compatibilty tools, and the
performance numbers for sci_copy (scibench2) came back up to
expected levels for the target system.
Since SCI allows userspace -> userspace copying across systems,
the problem code was being generated in user space and not kernel.
We did not attempt to track any further than installing just the
base RedHat 7.1 Fischer release -- the numbers were acceptable
for the target system. The performance numbers attached to this
release wer run on a PIII Intel System using the 440BX Bridge
chipset on a system with a maximum PCI throughput of 87 MB/S.
NOTES:
*Fixed concurrent DMA and PIO error detection bug for D330 adapters
*Linux:Fixed problems to allocate ATT tables for large SCI memsizes.
*Linux:Added proper handling of driver unload if driver load fails.
*Linux:fixed sleepOk flags to allow __GFP_WAIT flags to be passed to
__get_free_pages() to correct alloc failures in Linux when
memory gets fragmented.
*Linux:fixed ASSERT() macros for modversioned IRM builds on Linux
systems.
*Added SIMD/MMX support / Windows 2000. Not ported to Linux yet.
(coming to Linux soon)
*Added proper handling of SCIInitialize() failiure in all
examples/demos.
*UNIX: Fixed SCIMapLocal/RemoteSegment() bug when mapping to a segment
offset.
*Linux:added code to /src/adm/MAKE/CONFIG-LINUX to append the
-D__SMP__ build option when SMP support is detected in
kernel source tree to correct SMP build errors on 2.2/2.4 kernels JVM
Please note: call to SCIInitialize/SCITerminate will be mandatory from
next version of SISCI.
SISCI applications should update code ASAP. SCIInitialize()
must be the first SISCI function used. SCITerminate() should be the
last one used.
/opt/DIS/bin/scibench2 compiled Feb 7 2001 : 19:15:19
Test parameters for client
----------------------------
Local nodeId : 4
Remote nodeId : 8
Local adapter no. : 0
Segment size : 65536
Loops to execute : 1
ILoops to execute : 1000
Key Offset : 0
Source buffer type : User malloc
Memcopy mode : scicopy
Direction : PUSH
Error Check : No
SISCI API version : SISCI API version 1.10.0 ( Feb 7 2001 )
Adapter type : D320
Serial number : 100963
Hostbridge : 440BX
IO Bus frequency : 33 MHz
SCI Link frequency : 100 MHz
B-Link frequency : 66 MHz
----------------------------
Connect to remote segment (id=0x80400) .... connected
Remote segment (id=0x80400) is connected.
Remote segment (id=0x80400) is mapped to user space.
-- Starting the data transfer --
----------------------------------------------------
Segment Size: Average Transfer Time: Throughput:
----------------------------------------------------
4 2.14 us 1.87 MBytes/s
8 2.22 us 3.61 MBytes/s
16 2.53 us 6.32 MBytes/s
32 4.25 us 7.53 MBytes/s
64 3.42 us 18.70 MBytes/s
128 6.07 us 21.09 MBytes/s
256 6.34 us 40.40 MBytes/s
512 6.46 us 79.28 MBytes/s
1024 12.57 us 81.46 MBytes/s
2048 29.53 us 69.36 MBytes/s
4096 50.59 us 80.96 MBytes/s
8192 100.85 us 81.23 MBytes/s
16384 208.46 us 78.60 MBytes/s
32768 404.03 us 81.10 MBytes/s
65536 808.05 us 81.10 MBytes/s
Node 4 triggering interrupt
Interrupt message sent to remote node
The remote segment is unmapped
The segment is disconnected
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
Please read the FAQ at http://www.tux.org/lkml/
