Dear ROOTers!
Recently I have been tasked to put together new computer for data analysis, mostly in ROOT, for the new Tandetron laboratory. The data sets for the analysis is expected to be in order of a hundreds of GB up to a few TB. I would like to ask you for hardware recommendation for such computer, or what HW parameters should be chosen more delicately.
For all I know, the CPU should be oriented towards single thread performance, but what about CPU cache, RAM size/clock speed and GPU ?
Who will use these computers, what are their coding skills? Will they write the analysis themselves or will they use a framework? Are you planning on creating a cluster with a batch system or is this for interactive analysis?
In general, 1gb ram per core is a reasonable minimum for most analyses. SSDs are always nice for interactive use, of you can afford them. Maybe even distribute the files across the cluster and transfer then through the network, e.g. using http.
In general, ROOT and data analysis in general is fairly parasitic: it takes whatever it gets ROOT does have several concurrent modes, eg proof, schedulers like multiproc, or ROOT’s imt mode. But indeed, single core is always the simplest option…
We can use GPUs for deep neural nets (TMVA), CUDA=Nvidia helps here.
Hey Axel,
thanks for your respond, I should have been probably more specific. I imagine that the code will be very simple - going through trees using a few while cycles with conditions for single gamma and gamma-gamma coincidences. Thus I would say that the coding skill will be fairly low .
In some cases interactive analysis will be used also. Data will be obtained from NAS, internal RAID or the TCP stream of online data from acquisition server. PCIe SSD will be only used as a system disk.
The general idea is to have “normal” PC that will be used only for analysis and the question is if there are some key CPU/MB/RAM/… parameters that would save some computation time if chosen properly. There are no plans to employ proof, multiproc or GPU because of lack of experience with these modes
Then just get the fastest single core. Analysis is usually not that resource hungry if designed well, and ROOT sort of forces that on you Usually you can get a few percent improvement from better hardware, but factors from better analysis code. So please invest there
ROOT does have several concurrent modes, eg proof, schedulers like multiproc, or ROOT’s imt mode. But indeed, single core is always the simplest option…
Usually you can get a few percent improvement from better hardware, but factors from better analysis code. So please invest there