ALINKA Linux Clustering Letter:
abrenner at alinka.com
Wed Apr 12 12:32:13 MDT 2000
Alinka announces the weekly ALINKA Linux Clustering Letter:
clustering at alinka.com is a free weekly e-mail newsletter on linux clustering
To subscribe to the list, send e-mail to clustering at alinka.com from the
address you wish to subscribe, with the word "subscribe" in the subject.
To unsubscribe from the list, send e-mail to clustering at alinka.com from the
address you wish to unsubscribe from, with the word "unsubscribe" in the
Alinka is the editor of the ALINKA ORANGES and ALINKA RAISIN administration
software for Linux clusters. (Web site: http://www.alinka.com )
clustering at alinka.com provides a summary of the weekly activity in
relative to linux clustering (such as beowulf, linux virtual server or
general clustering news.
Here is the first Alinka Linux Clustering Letter:
This is the ALINKA Linux Clustering Letter of Wednesday, April the 12th.
News from the High Performance world, by Dr Laurent Gatineau
(lgatineau at alinka.com)
New Beowulf cluster in the world
The University of New Mexico  and IBM  have made a cluster of
256 IBM Netfinity servers (dual processors). We could read in
Wired  that this cluster will deliver a processing speed of 375
gigaflops, or 375 billion operations per second. It will only rank
24th on the list of the top 500 fastest supercomputers .
It's clear that we need good software and good hardware before that
clusters could be better than supercomputer. As Dr. Frank Gilfeather
said, Beowulf clusters need management tools like on supercomputer,
and they need good I/O, including scalable file systems.
The Forecast Systems Laboratory  has made a Beowulf cluster for
numerical weather prediction. We could read in the Linux Weekly
News  that the FSL cluster (called "Jet") currently consists of 276
nodes. Each node is a 667Mhz Alpha processors with 512 Mb of
memory. Like for the LosLobos cluster (and lots of Beowulf cluster),
the interconnection network used is Myrinet.
The University of Maine  has made a Beowulf cluster of 34
nodes. Each node is based on two 600Mhz Pentium III with 512 Mb of
memory and two Fast Ethernet cards. This cluster is available for
testing, computational physics and computational physical chemistry.
Conference about Clustering
Extreme Linux Workshop/Conference #3
EL2000 will be held in conjunction with the 4th Annual Linux
Showcase & Conference on October 12-14, 2000 in Atlanta, GA. Besides
the Extreme Linux track there will be two other refereed tracks: Hack
Linux and Use Linux. There will also be vendor exhibits, tutorials,
birds of feather sessions, and work in progress sessions. Attendees
of the Extreme Linux workshop will be full attendees of the conference
and able to attend any sessions they want.
Software for Beowulf cluster
dsh - distributed shell
dsh executes one or more commands on a collection of hosts. The hosts
may be specified on the command line or as nodegroups. Commands are
executed sequentially on each host, and the output from each command
is prepended with the hostname. If a command is not specified on the
command line, the user is prompted for commands to execute.
From: Beowulf mailing list
Home Page: http://www.ccr.buffalo.edu/dsh.htm
PVFS Kernel Interface v0.8.1
Allows mounting of PVFS file systems on Linux machines running 2.2.xx
kernels (tested on 2.2.12, 2.2.13, 2.2.15pre4). This version no longer
requires patching the kernel.
Home Page: http://www.parl.clemson.edu/pvfs
SCMS 1.2 Early Test Version
Early version SCMS1.2 beta is now available for testing. If you try
and it doesn't work , please report the problem. New feature
1. All in Java now. Much better user interface
2. KCAP is now a separate package with many improvement.
3. Realtime monitoring work much better.
See [m1] and [m2] for downloading
From: the beowulf mailing list
TCP patches for Red Hat 6.2
Josip Loncaric's TCP patches for Red Hat 6.2 (Linux kernel 2.2.14-5.0)
are available. See:
Patch for RH6.2: http://www.icase.edu/~josip/tcp-patch-for-2.2.14-5.0
From: the beowulf mailing list
Tips and tricks from the Beowulf mailing list
* There was some threads about memory and cluster: how to buy / test /
bench it ?
A simple conclusion could be: "buy expensive memory and you won't
have problems", but it's not always true and Beowulf clusters should
use all kind of hardware, so if you could trust in your vendor buy
its memory and test it. In fact you should always test your hardware
before using it.
To test it there is some tools:
. the famous memtest86 
. Adam Lazur [m1] gives the URL of memtester  and said that "he
prefers memtester over memtest86 as it has a lot of algo's for
finding bad RAM"
Another thing about the memory tester, Douglas Eadline [m2] reports
that "floating point on x86 hits the RAM espacially hards", he
hasn't verified this.
An other way to test a node is to make a benchmark of the memory. A
good tool for this is stream . By testing the performance of your
memory you test also important hardware such as the cache and the
bus, and you could test the scalability of SMP nodes. This is the
subject of the David Konerding's thread.
* Ole Holm Nielsen had written a Beowulf cluster mini-HowTo, you could
find it at this url . Documentations are important: thanks for
this work !
* Borries Demeler asks for a node cloning software... This is one of
the most important tools for a cluster management software (and its
fully provides in our software Alinka Raisin). Robert G. Brown [m2]
proposes the kickstart installation process from RedHat ; dwight
[m3] and Pfenniger Daniel [m4] find kickstart very usefull. Alex
Lancaster [m5] wrote that one of the limitation of kickstart is that
it's not easy to configure software that require human interaction.
* Alexander Korenkov [m1] seeks for some tricks to tune his MPI: with
his 100 Mbit fast Ethernet his net speed is only about 7.2 Mbit
/s. According to the description of his algorithm, the best answer
should be the Jeff Squyres's one [m2]: "use the persistent mode
sends and receives". He gives a good description on the cost to send
a message: you have to take into account the cost to go through the
system and the cost to go through the network (the first one could
be very expensive when you send lots of small messages).
* Jose Marin [m1] asks for a network traffic monitor. There's lots
of network meter, some does lots of thing, other are
lightweight... Robert G. Brown [m2] had wrote procstatd  and give
a pointer [m3] to mgm  which is probably not for professional
using. Jose Marin answers himself and gives a pointer to an SGI's
tool (Performance Co-Pilot)  which monitor all system ressources,
and two other pointers ,  for the tools iptraf and ntop which
are recommended by Jay Sherman, Felix Rauch [m4]. Two others tools
were cited by Lyle Bickley and Paul Nowoczynski: ethertape  and
Ethereal  (which is a network protocol analyzer).
* Kragen Sitaker [m1] wants to change the bios of all his nodes
without doing it by hand. Erik Arjan Hendriks [m2] gives a solution
with the tool Bios Writer .
* J.Dube [m1] wants to know if "a Beowulf on an outside connection is
a huge security hole". A short answer could be that a Beowulf is
like one computer, so it's no more no less a security hole than to
put a computer on an outside connection... It depends on how the
Beowulf is configured. Robert G. Brown [m2] wrote a long mail
explaining how the Beowulf could be configured to close the maximum
of security holes, and dwight [m3] added some comments.
News from the High Availability world, by Christophe Massiot
(cmassiot at alinka.com)
* Jean-Christophe Boggio describes [lvs1] a solution for having load
balancing with MySQL. The application must separate read queries from
write queries, read queries are dispatched normally through lvs, and
write queries are done as many times as there are SQL servers at the
application level. Sean Ward has developed such a solution which reads
log files for replication [lvs2]. The source code is currently
* Phil Z. sends in a report [lvs4] on Realnetworks G2 server. To have
it work under LVS, the audio/video daemon must be configured to
listen/respond to both the VIP and its real IP, and both port 7070
and 554 must be redirected.
* Using LVS for balancing smtp servers should work, provided you do not
have ident loops with multiple smtp servers. To avoid this, turn on the
-R option of tcpserver (qmail/tcpserver) [lvs5], or read the HOWTO for
sendmail [lvs6]. POP will only work if you have an NFS-safe POP server
* RedHat creates a new mailing-list for piranha, a web interface for
Linux clustering. [lvs8]
* IPVS patch 0.9.10 is out. Here is the ChangeLog :
* Julian added the droprate and secure_tcp defense strategies.
* The dropentry defense strategy was revisited.
* The fwmark service lookup was added by Horms, Julian and Wensong
Use a firewall-marking to denote a virtual service instead of a
triplet <protocol,addr,port>. The marking of packets with a
firewall-mark is done by firewalling code. This feature can be
used to build a virtual service assoicated to different IP
addresses or port numbers, but sharing the same real servers, such
as multiple-homed LVS. [lvs9]
* A document on LVS defense strategies against DOS attack is available.
* Legato has released a clustering solution called Legato Cluster
Enterprise. It supports Solaris, Linux (RedHat, Caldera), Windows NT/2000,
HP-UX and AIX platforms. [lha1]
* Mike Wangsmo has released kernels with ext3 patches for testing purposes
* Chris Wright summarizes the status of LinuxFailSafe [lfs1]. The port
to Linux is well under way, and being done largerly by SGI (release date
in the summer ?). The intent is to open source almost all of FailSafe,
with a licence near GPL/LGPL.
* The presentations made by SGI at the Linux FailSafe Symposium at Denver
on 31st March are now available. [lfs2]
News on the Filesystems front, by Ludovic Ishiomin
(lishiomin at alinka.com)
In [1m], someone asks if intermezzo could be used for filesystem
replication, in order to achieve high availability. The answer
is that intermezzo can handle this job, but it's not ready
yet for critical tasks.
JFS 0.0.5 for Linux, a well known journaled filesystem which
comes from IBM AIX, has been announced. More information can be
found at . But it is not yet ready for a public release.
In the linux-lvm mailing list, there had been a discussion about the
ability to convert an existing filesystem into a logical volume.
The conclusion is that it should be possible, but some code is
needed in the ext2resize utility.
This letter was brought to you by ALINKA (http://www.alinka.com), the
editor of the ALINKA ORANGES
and ALINKA RAISIN administration software for Linux clusters.
abrenner at alinka.com
http://www.alinka.com; AlinkA : Cluster Solutions
Tel: (+33) 1 49 35 29 29
More information about the Linux-HA