Re:Lets get this guy to back up his words =)
on
Macintosh Clustering
·
· Score: 1
Last October, when Dr. Decyk and I presented at the Aerospace Corporation, we accidentally showed up late. There were a bunch of powered-off PBG3's in bags there we had just seen for the first time. We brought one Asante Ethernet switch and some cables.
10 minutes later, in front of the audience, we had those PB's working as a cluster, and we were ready to present. We were told the audience was impressed. I could forward you the emails of some of the Aerospace employees who witnessed us.
You are certainly welcome to try out the software for yourself.
Have fun,
Dean
P. S. As for getting work done, you may critique my dissertation at: http://dauger.com/DaugerDissertation.pdf All the data was computed, analyzed, and visualized on Macs.
I'm afraid the writer did misstate me. It was "How to Build a Beowulf" by Sterling, et al.
The first AppleSeed-type cluster was created by researchers in 1998 at the Plasma Physics Group at UCLA:
http://exodus.physics.ucla.edu/appleseed/
You may use the Other Clusters link to find out about Mac clusters in other research settings.
The software works over the Internet. Using the Remote Scan feature in Pooch, I've run a parallel job combining the UCLA Mac cluster with my Mac while I was 6000 miles away in Garching (outside Munich), Germany. (Obviously, the latency was pretty bad, but it worked.) Please try it and see for yourself.
The achieved performance of one dual processor G4/533 is 3.6-3.9 GF with the AltiVec Fractal Carbon demo, depending on the Max Count settings. I recommend going to Taper Hall and try it out yourself. We added about 20 dual-processor G4/450s elsewhere on campus to those 56.
Please define your "cost" and include it in your "price". How much is your time worth? Multiply that by the time you will have to spend maintaining your cluster.
The Plasma Physics group at UCLA does not have the money to pay for Linux expertise. That's why we went with the Mac to get our work done, which you can see a sample of at:
When we walked away, the nodes were still ready for the undergraduates for use as a computer lab. This cluster software coexists with mainstream apps, like Office.
Not all scientists are skilled in computers to the degree you are talking about. It's a wide range. There are many that will attest to that besides me.
If you would like to find out more about others using Mac clusters, have a look at these links:
http://exodus.physics.ucla.edu/appleseed/applese ed sites.html
http://exodus.physics.ucla.edu/appleseed/HiSchoo l/ HiSchool.html
http://daugerresearch.com/pooch/users.html
http://exodus.physics.ucla.edu/appleseed/dev/dev el oper.html
We're using MPI jobs and codes that ran on the Cray's, IBM SPs, Fujitsu's, and SGI's.
Dean
Re:No wonder he doesn't get consulting jobs...
on
Macintosh Clustering
·
· Score: 1
Talk to Donald Becker of Scyld and co-author of "How to Build a Beowulf". Talk to Office of Academic Computing at UCLA. Talk to the High-Performance Computing Group at JPL.
The achieved performance of one dual-processor G4/533 was 3920 MF. There were 56 dual-processor G4/533's plus 20 dual-processor G4/450's. You can see some of the machines we used on the top floor of Taper Hall at USC.
As for implementation details please read this:
http://exodus.physics.ucla.edu/appleseed/applese ed.html#report
http://exodus.physics.ucla.edu/appleseed/dev/dev el oper.html
The Fractal demos, and the scientific codes that run on AppleSeed clusters, use MPI. For more info and source code:
http://exodus.physics.ucla.edu/appleseed/dev/dev el oper.html
>> go around with a CD and reboot every computer
You could set the Software Update to automatically update. There is a Macintosh Manager utility for more detailed adminstration. And Pooch can be upgraded by launching a Pooch Package in parallel in cluster, updating the clustering software in seconds. (Try launching the Pooch Installer in parallel on a Pooch cluster.)
Dean
Re:Why the Mac won't be a good clustering choice
on
Macintosh Clustering
·
· Score: 1
With no extra hardware purchase, we turned a idle, unused USC computer lab into a 233-GF cluster. When we were done, it was still a computer lab, ready for the undergraduates. I think that's very economical.
Only Pooches of the same registration can talk to each other. It's like sharing access to a common office; the security depends on how well the users keep their keys.
At the IEEE Cluster 2001 conference, one of the authors of Scyld and "How to Build a Beowulf" said that the version skew between kernels on the cluster is an issue that can break applications. This is also known to occur at the Linux Beowulfs at UCLA, so once they have it running they say "don't touch it!" because it's so fragile. The Scyld solution was to customize Linux purely for the purpose of clustering to the exclusion of almost all other software.
Pooch and MacMPI_X can run on both OS X 10.1, a Unix, and OS 9, not a Unix, including a cluster consisting of a mix of the two. Try it yourself. That's got to be the biggest version skew between OS's I've heard of. And the software easily coexists with mainstream apps, such as Microsoft Office.
>> Is it assumed that apple seeds will not be connected to the internet in any way, nor have any wireless access point attached to them?
Commands between Pooches are protected by an internal, rotating 512-bit encryption key based on the registration name of the user. Therefore, Pooches only of the same registration name can talk to each other. The idea is like shared access to an office resource. The demo version is relatively insecure because, well, anybody can download it. For more info, see the dox at:
http://daugerresearch.com/pooch/download.html
There are plenty of AppleSeed clusters connected to the Internet. I've accessed the one at UCLA from lots of places: using my PBG4 via Airport from my couch in Pasadena, from Toronto, Canada, from Garching, Germany (outside Munich)....
There is someone who charges $250/client for Linux on Intel Beowulf-type software.
Slashdot Mirror
Last October, when Dr. Decyk and I presented at the Aerospace Corporation, we accidentally showed up late. There were a bunch of powered-off PBG3's in bags there we had just seen for the first time. We brought one Asante Ethernet switch and some cables.
10 minutes later, in front of the audience, we had those PB's working as a cluster, and we were ready to present. We were told the audience was impressed. I could forward you the emails of some of the Aerospace employees who witnessed us.
You are certainly welcome to try out the software for yourself.
Have fun,
Dean
P. S. As for getting work done, you may critique my dissertation at: http://dauger.com/DaugerDissertation.pdf All the data was computed, analyzed, and visualized on Macs.
I'm afraid the writer did misstate me. It was "How to Build a Beowulf" by Sterling, et al.
The first AppleSeed-type cluster was created by researchers in 1998 at the Plasma Physics Group at UCLA:
http://exodus.physics.ucla.edu/appleseed/
You may use the Other Clusters link to find out about Mac clusters in other research settings.
The software works over the Internet. Using the Remote Scan feature in Pooch, I've run a parallel job combining the UCLA Mac cluster with my Mac while I was 6000 miles away in Garching (outside Munich), Germany. (Obviously, the latency was pretty bad, but it worked.) Please try it and see for yourself.
Have fun,
Dean
>> What am I missing?
The achieved performance of one dual processor G4/533 is 3.6-3.9 GF with the AltiVec Fractal Carbon demo, depending on the Max Count settings. I recommend going to Taper Hall and try it out yourself. We added about 20 dual-processor G4/450s elsewhere on campus to those 56.
Have fun,
Dean
Please define your "cost" and include it in your "price". How much is your time worth? Multiply that by the time you will have to spend maintaining your cluster.
e .h tml
e r/ USCMacClusterBenchmark.html
The Plasma Physics group at UCLA does not have the money to pay for Linux expertise. That's why we went with the Mac to get our work done, which you can see a sample of at:
http://exodus.physics.ucla.edu/appleseed/scienc
Also, we took an idle, unused USC computer lab and used it as a cluster with zero additional hardware and achieved these results:
http://daugerresearch.com/fractaldemos/USCClust
When we walked away, the nodes were still ready for the undergraduates for use as a computer lab. This cluster software coexists with mainstream apps, like Office.
Have fun,
Dean
Not all scientists are skilled in computers to the degree you are talking about. It's a wide range. There are many that will attest to that besides me.
e ed sites.html
o l/ HiSchool.html
If you would like to find out more about others using Mac clusters, have a look at these links:
http://exodus.physics.ucla.edu/appleseed/apples
http://exodus.physics.ucla.edu/appleseed/HiScho
http://daugerresearch.com/pooch/users.html
Have fun,
Dean
For more documentation please see:
v el oper.html
e ed .html#report
http://exodus.physics.ucla.edu/appleseed/dev/de
http://daugerresearch.com/pooch/download.html
http://exodus.physics.ucla.edu/appleseed/apples
Have fun,
Dean
OS 9 and OS X versions of the Pooch Quick Start and Manual are at:
http://daugerresearch.com/pooch/download.html
I just didn't want to squeeze that many screen shots into one page. The software works on both OS X 10.1 and OS 9. Please try it yourself.
Have fun,
Dean
>> This artilce is just pure Mac FUD the more I think about it.
e ed .html
Please read this web site and the AppleSeed report:
http://exodus.physics.ucla.edu/appleseed/apples
Dean
Have a look at:
v el oper.html
http://exodus.physics.ucla.edu/appleseed/dev/de
We're using MPI jobs and codes that ran on the Cray's, IBM SPs, Fujitsu's, and SGI's.
Dean
Talk to Donald Becker of Scyld and co-author of "How to Build a Beowulf". Talk to Office of Academic Computing at UCLA. Talk to the High-Performance Computing Group at JPL.
Thank you so much for your objective language.
Dean
Please look at:
e r/ USCMacClusterBenchmark.html
e ed .html#report
v el oper.html
http://daugerresearch.com/fractaldemos/USCClust
The achieved performance of one dual-processor G4/533 was 3920 MF. There were 56 dual-processor G4/533's plus 20 dual-processor G4/450's. You can see some of the machines we used on the top floor of Taper Hall at USC.
As for implementation details please read this:
http://exodus.physics.ucla.edu/appleseed/apples
http://exodus.physics.ucla.edu/appleseed/dev/de
Dean
>> it would be nice to see some bench marks
e r/ USCMacClusterBenchmark.html
a rk s.html
Look at:
http://daugerresearch.com/fractaldemos/USCClust
http://exodus.physics.ucla.edu/appleseed/benchm
Dean
>> standard communication libraries
v el oper.html
The Fractal demos, and the scientific codes that run on AppleSeed clusters, use MPI. For more info and source code:
http://exodus.physics.ucla.edu/appleseed/dev/de
>> go around with a CD and reboot every computer
You could set the Software Update to automatically update. There is a Macintosh Manager utility for more detailed adminstration. And Pooch can be upgraded by launching a Pooch Package in parallel in cluster, updating the clustering software in seconds. (Try launching the Pooch Installer in parallel on a Pooch cluster.)
Dean
With no extra hardware purchase, we turned a idle, unused USC computer lab into a 233-GF cluster. When we were done, it was still a computer lab, ready for the undergraduates. I think that's very economical.
Dean
>> But as usual, it looks like the security ramifications for this are enormous,
p df
Have a look at the Security chapter of the documentation at:
http://daugerresearch.com/pooch/PoochManualX.1.
Only Pooches of the same registration can talk to each other. It's like sharing access to a common office; the security depends on how well the users keep their keys.
Dean
My philosophy is that the best programmers are the ones who make every effort to make themselves obsolete.
Dean
At the IEEE Cluster 2001 conference, one of the authors of Scyld and "How to Build a Beowulf" said that the version skew between kernels on the cluster is an issue that can break applications. This is also known to occur at the Linux Beowulfs at UCLA, so once they have it running they say "don't touch it!" because it's so fragile. The Scyld solution was to customize Linux purely for the purpose of clustering to the exclusion of almost all other software.
Pooch and MacMPI_X can run on both OS X 10.1, a Unix, and OS 9, not a Unix, including a cluster consisting of a mix of the two. Try it yourself. That's got to be the biggest version skew between OS's I've heard of. And the software easily coexists with mainstream apps, such as Microsoft Office.
Have fun,
Dean
Have you tried it?
Dean
>> Is it assumed that apple seeds will not be connected to the internet in any way, nor have any wireless access point attached to them?
Commands between Pooches are protected by an internal, rotating 512-bit encryption key based on the registration name of the user. Therefore, Pooches only of the same registration name can talk to each other. The idea is like shared access to an office resource. The demo version is relatively insecure because, well, anybody can download it. For more info, see the dox at:
http://daugerresearch.com/pooch/download.html
There are plenty of AppleSeed clusters connected to the Internet. I've accessed the one at UCLA from lots of places: using my PBG4 via Airport from my couch in Pasadena, from Toronto, Canada, from Garching, Germany (outside Munich)....
There is someone who charges $250/client for Linux on Intel Beowulf-type software.
Dean
> detailed work exists
e ed .html#report
e .h tml
v el oper.html
Yes, detailed work does exist regarding Pooch and in the history of AppleSeed. Look at:
http://exodus.physics.ucla.edu/appleseed/apples
http://exodus.physics.ucla.edu/appleseed/scienc
For some documentation see:
http://exodus.physics.ucla.edu/appleseed/dev/de
http://daugerresearch.com/pooch/download.html
Have fun,
Dean