I just wanted to show a sign of life after staying away for approx a year and share how my HW had been used since I left the FAH crowd in these 12 months.
Reading through some of the threads the departure from the FAH project seems to be a wider spread movement - showing the complexity and effort required to run a large dispersed community for a longer period.
With all the HW investments for folding and sorting still available, some of the systems got reconfigured for different objectives - supporting students in local universities in their quest for computational resources to conduct research based on simulations.
While I enjoyed the work in distributed computing for a while, I found this second phase more rewarding. By not only donating my systems for research done by students I had the pleasure to meet, those projects were insightful and fun as well.
Take for instance a student working in bio informatics utilizing Gromacs or Rosetta to simulate proteins for a particular research interest or exam. Gromacs has a significantly different CPU/GPU computational pattern than FAH and so the FAH optimized machines were quite sub-optimal for Gromacs code. Working with students to identify barriers or bottlenecks , reconfiguring systems, optimizing code bases to different GPU microarchitecures, tuning HW appropriately is such a fun past time, that I might stick for the next year or so with this approach, until my HW arsenal ages out.
I appreciate and value all of you what you are doing to support distributed compute projects around the globe, but wanted to share this opportunity to support also local research and local students as well. It is real fun ...
Some of the experiences:
So the one 4-socket server, the three 2-socket servers and the 7 dualGPU cubes are individually or in subsets "distributed" to a group of different people I trust, or run (more often) in my house. Anyway, it is fun and great to see how ubiqutous simulation has become in many scientific fields.
I learned a lot here @Hardforum and enjoy remembering those fond days which helps me in this new phase.
All the best to you guys (and gals),
Andy
Reading through some of the threads the departure from the FAH project seems to be a wider spread movement - showing the complexity and effort required to run a large dispersed community for a longer period.
With all the HW investments for folding and sorting still available, some of the systems got reconfigured for different objectives - supporting students in local universities in their quest for computational resources to conduct research based on simulations.
While I enjoyed the work in distributed computing for a while, I found this second phase more rewarding. By not only donating my systems for research done by students I had the pleasure to meet, those projects were insightful and fun as well.
Take for instance a student working in bio informatics utilizing Gromacs or Rosetta to simulate proteins for a particular research interest or exam. Gromacs has a significantly different CPU/GPU computational pattern than FAH and so the FAH optimized machines were quite sub-optimal for Gromacs code. Working with students to identify barriers or bottlenecks , reconfiguring systems, optimizing code bases to different GPU microarchitecures, tuning HW appropriately is such a fun past time, that I might stick for the next year or so with this approach, until my HW arsenal ages out.
I appreciate and value all of you what you are doing to support distributed compute projects around the globe, but wanted to share this opportunity to support also local research and local students as well. It is real fun ...
Some of the experiences:
- How badly Gromacs scale to run efficiently on 4-socket machines (it does well on 2-socket servers), due to lack of reasonable NUMA optimization
- Rosetta "loves" Intel CPUs, but lacks NUMA optimization as well
- How much more mature the CUDA ecosystem and programming model currently is vs. OpenCL.
- How little scientists and students considered computational intensity in home grown algorithms (none used watt meters ...)
- The incredible efficiency and performance of the Maxwell GPU in Gromacs (BTW, will pick up my EVGA overclocked GTX 980 tomorrow)
- How universally balanced the relatively old Radeon 7970 GE is for a range of OpenCL codes.
- That quite a few codes run almost at the same speed on the 4-core 4790K and 6-core 3930K (and how hot the 4790K gets)
- The convenience, utility and stability of running dual-socket water-cooled Xeon systems in the Coolermaster HAF BX cube including 2 GPUs (with small ATX Asus motherboards)
- The importance of system stability. Some of the codes used run for 7-10 days with no restart functionality built in. HW better be stable.
So the one 4-socket server, the three 2-socket servers and the 7 dualGPU cubes are individually or in subsets "distributed" to a group of different people I trust, or run (more often) in my house. Anyway, it is fun and great to see how ubiqutous simulation has become in many scientific fields.
I learned a lot here @Hardforum and enjoy remembering those fond days which helps me in this new phase.
All the best to you guys (and gals),
Andy