[Gamma_sw] A new version of GAMMA... (fwd)
Eugene.Leitl at lrz.uni-muenchen.de
Thu Jul 5 03:39:35 PDT 2001
-- Eugen* Leitl <a href="http://www.lrz.de/~ui22204/">leitl</a>
ICBMTO : N48 10'07'' E011 33'53'' http://www.lrz.de/~ui22204
57F9CFD3: ED90 0433 EB74 E4A9 537F CFF5 86E7 629B 57F9 CFD3
---------- Forwarded message ----------
Date: Wed, 4 Jul 2001 18:51:58 +0200 (MET DST)
From: Giuseppe Ciaccio <ciaccio at disi.unige.it>
To: GAMMA mailing list <gamma_sw at lists.dsi.uniroma1.it>
Cc: makino at astron.s.u-tokyo.ac.jp, Dan.Lizotte at unb.ca,
Winfrid.Tschiedel at fujitsu-siemens.com, Cosimo Anglano <mino at unipmn.it>,
Marco Danelutto <marcod at di.unipi.it>,
Vincenzo Di Martino <vincenzo at caspur.it>,
Philippe Augerat <Philippe.Augerat at imag.fr>, Korah <korahs at vsnl.com>
Subject: [Gamma_sw] A new version of GAMMA...
...has been released just a few minutes ago (at last!).
This release marks a substantial improvement over the past ones.
In a few words:
1) Support for mixed (IP and GAMMA) traffic on the same network has been
tested and debugged, and now is hopefully stable (a number of bugs
were affecting the GAMMA driver for the AceNIC adapter).
2) Support for Intel EtherExpress Pro/100 and Packet Engines GNIC-II NICs
has been discontinued, due to shortage of NICs to play with and lack
3) Blocking send routines of GAMMA now enjoy a higher degree of reliability,
thanks to a mechanism for detecting and retransmitting missing packets.
This is of great help, especially for large clusters (where the
probability of a frame error get higher) and generic parallel jobs
(usually characterized by bad communication patterns).
4) The installation manual has been updated.
A new version of MPI/GAMMA has been released as well, as the use of GAMMA
non-blocking send routines in MPI/GAMMA has been temporarily disabled
(it will be re-enabled as soon as the GAMMA packet retransmission mechanism
will be extended to the non-blocking case).
Go browse the GAMMA and MPI/GAMMA web pages,
and enjoy! Comments and reports (of bugs, of course...) are welcome.
This release of GAMMA has been tested on a cluster counting four to six
nodes (courtesy of INFN, sezione di Genova), each equipped with three NICs:
a Netgear GA620 (Gigabit Ethernet), a DEC DE500 (Fast Ethernet), and a
3COM 3c905C (Fast Ethernet). This made it possible to easily test all
the GAMMA drivers using a single testbed.
Tests include the NAS Parallel Benchmarks (class A) and the "ktkmpi"
switch-congesting MPI test (courtesy of Massimo Bernaschi).
Packet retransmission will be eventually extended to cover the non-blocking
case (although there will be inherent limitations). The current implementation
is based on negative acks and timeouts on receive, so as not to require
any additional overhead in the "usual case" (in my opinion, the
"usual case" is when no congestion occurs and the probability of frame error
is negligible...should I rather call this "the lucky case"?). Any request
for packet retransmission is interpreted as a sign of congestion hazard in the
network; in an attempt to "react" to the hazard, the credit size of
the flow control mechanism is decreased at those nodes who get a retransmission
request (and then is gradually increased if no imore packet losses occur).
The effectiveness of such "reaction" will be investigated.
Giuseppe Ciaccio http://www.disi.unige.it/person/CiaccioG/
DISI - Universita' di Genova via Dodecaneso 35 16146 Genova, Italy
phone +39 10 353 6638 fax +39 010 3536699 ciaccio at disi.unige.it
gamma_sw mailing list
gamma_sw at lists.dsi.uniroma1.it
More information about the Beowulf