IGT HPC Work Group Meeting - Scalable x86-based solutions for HPC

-

IGT HPC Work Group Meeting - Scalable x86-based solutions for HPC Users: Clusters, Cloud, and Shared Memory (SMP) in a unified platform


Wednesday, March 3rd, 2010
IGT Offices, Maskit 4, 5th Floor, Hertzliya.
13:45-14:00: Gathering
14:00-14:05: Opening - Avner & Guy
14:05-15:00: Part 1
15:00-15:10: Coffee Break
15:10-16:00: Part 2
16:00-16:15: Discussion and concluding remarks

Title:Scalable x86-based solutions for HPC Users: Clusters, Cloud, and Shared Memory (SMP) in a unified platform
Abstracts:IBM (Gilad Berman, System x Technical Specialist)Scalable Data Center alternatives: iDataplex, BladeCenter, new management softwareiDataPlex technology offers a new innovative design to address the needs of the data center with extreme density and simplified manageability while reducing power and cooling consumption.BladeCenter is an established, open design with a flexible, scalable architecture that offers a true alternative to today’s sprawling racks and overheated server rooms. By integrating servers, storage, networking and management, BladeCenter is helping companies maximize performance, minimize costs and sweep complexity aside.xCAT is a new advanced generation of cluster deployment and management system.
ScaleMP (Benzi Galili, Executive Vice President)SMP machines (shared memory, symmetric multi-processor) are still required for several classes of HPC workloads, mainly for use of very large (>0.5TB) RAM and also for high-core counts. In the past, building SMPs required tens to hundreds of millions of ongoing project investments for custom-chipsets, and in many cases involved custom-everything (CPUs, interconnect, chipset, OS, and software stacks) – making this type of technology cost-prohibitive for many science and engineering organizations. Past commercial attempts at creating software-based SMP (a.k.a Software-DSM) have failed. ScaleMP uses a novel virtualization-based approach, taking advantage of high bandwidth commodity interconnect and high-performance standard (x86) processors, to create SMPs that perform better for HPC workloads, and cost significantly less than the hardware-only alternatives. This session will describe the types of workloads that benefit from the architecture (including use cases around the world), and some of the basic ideas behind the architecture.
Three usage classes will be discussed: classic SMP, HPC cluster simplification, and Cloud/Grid oriented SMP “on-demand” services.
16:25-16:35: Discussion and concluding remarks
 To register, please send your contact details to: info@grid.org.il

We are looking forward to seeing you!
Guy Tel-Zur, Ph.D. IGT Chairmanwww.grid.org.il

Comments

Post a Comment

Popular posts from this blog

The End of Grid Computing?

-
Image
In the year 2003 the MIT Technology review ranked "Grid Computing" among the 10 Emerging Technologies That Will Change the World [1]. We are now four years later and something is not going well with "Grid Computing". An indication that there is a problem can easily be seen by looking at the "Google Trends" plot for the term "Grid Computing": (click on the image to get the current trend). This finding can be compared with another buzz word, "Virtualization", which is older than "Grid Computing" and yet is gaining more and more momentum: There is however one exception. The Academic Grid is still having lot's of glory thanks to the huge heavily funded European (EGEE) and other US projects. When LHC data will start to be taken at CERN it will reach it's top importance. But, it seems that for other scientific projects Grid Computing is not going to be such a success. It will remain as "Nice to have" but will never...
Read more

Parallel Debugging with a Serial Debugger

-
Image
Parallel Debuggers are important tools when trying to eliminate bugs in parallel programs. Commercial parallel debuggers are expensive. In this blog post I will show you a free and open source alternative which is good enough for education purposes and for small codes. Suppose you want to debug the famous cpi.c code which computes an approximation to pi by a parallel numerical integration. This demo uses the free Data Display Debugger (ddd) which is based on gdb but has a nice GUI. Step 1, compile: mpicc -g -o cpi ./cpi.c Step 2, run (without a debugger): mpirun -np 2 ./cpi Step 3, run with the free Data Display Debugger (ddd) : mpirun -np 2 ddd ./cpi The last command will open two MPI processes (instances) of ddd each running cpi, see screen capture: Finally, I enclose here a short video which I hope is convincing: hi-res version:  https://youtu.be/N_J4NKJkuMs Happy (Free) MPI debugging!
Read more

[mpich-discuss] Errors while testing MPICH2 on Vista

-
[mpich-discuss] Errors while testing MPICH2 on Vista Guy Tel-Zur gtelzur at gmail.com Sat Oct 24 14:39:26 CDT 2009 Previous message: [mpich-discuss] MPI-3 Fortran feedback Next message: [mpich-discuss] MPI function to test a machine alive or not. Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] I am trying to test a fresh installation of MPICH2 (version 1.2) on Vista32 and get the following errors: c:\temp>"c:\Program Files\MPICH2\bin\mpiexec.exe" -n 4 cpi.exe Please specify an authentication passphrase for smpd: passphrase, e.g. behappy*> Error while connecting to host, No connection could be made because the target machine actively refused it. (10061) Connect on sock (host=user-PC, port=8676) failed, exhaused all end points Unable to connect to 'user-PC:8676', sock error: Error = -1 I will appreciate any comment. --Guy -------------- next...
Read more