My model became too complex and heavy to run on my desktop. I am considering now using the supercomputer resources at my university to run my simulations. some information I need to know:
1- As a MSWin user I have to switch now to UNIX to run my simulations on the supercomputer. Do I need to make any change to my old code that runs under the WINDOWS version of Neuron to run under the UNIX version of Neuron?
2- My model is a classified as an anatomically detailed model of a single cell. From discussions over the forum I know that NEURON is not currently prepared to to run over multiple CPUs and this feature is limited to simulations of networks. When this option will be available to single neuron models?
3- What is more important for NEURON to run a simulation faster: bigger memory or faster processor? Does the answer depend on the model itself?
4- At my university the machines are SGI MIPS processors running IRIX. Do you know of any problems installing Neuron under this version?
Thank you.
Best Regards,
Moving to a Parallel Hardware
1 - No. Not between mswin and unix. But the switch from serial to parallel does require thoughtful transformation.
2 - that's basically true but if you are willing to live on the bleeding edge and configure with --with-paranrn and partition the cell manually on the cpus and connect the partitions with gap junctions that just happen to have the right axial resistance and either use the fixed step method exclusively or (if using the global variable step method) avoid any use of NetCon then it can be done with the latest alpha version. By the end of the winter I hope to be able to separate the partitioning choice from the biophysical specification but it is kind of shaky whether I will meet that goal.
3- the answer depends on the model. but in this day of processors much faster than memory, the bigger the cache the better. You can often get superlinear speedup with parallel simulation due to being able to fit the entire model into high speed cache, ie. on the order of several hundred equations per cpu.
4- There is a good chance configure will succeed and no tweaks to the code will be needed. NEURON has run on many versions of IRIX in the past. If trouble arises I can help.
2 - that's basically true but if you are willing to live on the bleeding edge and configure with --with-paranrn and partition the cell manually on the cpus and connect the partitions with gap junctions that just happen to have the right axial resistance and either use the fixed step method exclusively or (if using the global variable step method) avoid any use of NetCon then it can be done with the latest alpha version. By the end of the winter I hope to be able to separate the partitioning choice from the biophysical specification but it is kind of shaky whether I will meet that goal.
3- the answer depends on the model. but in this day of processors much faster than memory, the bigger the cache the better. You can often get superlinear speedup with parallel simulation due to being able to fit the entire model into high speed cache, ie. on the order of several hundred equations per cpu.
4- There is a good chance configure will succeed and no tweaks to the code will be needed. NEURON has run on many versions of IRIX in the past. If trouble arises I can help.