How to profile in CUDA application with compute capability 7.x? Is metric "dram_read_throughput" valid in Nsight Compute?

Question

My setup environment: CUDA 10.2 Device: RTX 2080 OS: Ubuntu 16.04 When I try to use nvprof, I find that it doesn't support devices with compute capability 7.2 and higher. It is recommended that I should use Nsight Compute or Nsight Systems instead. But I can not launch the above two software because of the lack of graphical interface. How could I use Nsight Compute in remote server? By the way, is it possible to profile metrics in Nsight Compute?

Answer 1

How to profile in CUDA application with compute capability 7.x?

For compute capability 7.5 and higher the recommended tools are nsight compute, and nsight systems. The documentation for nsight compute is here, the documentation for nsight systems is here. There is an introductory blog describing these "new" CUDA profiler tools here, and a tutorial blog on nsight systems here and a tutorial blog on nsight compute here. The introductory blog describes why there are 2 tools, and how they relate to each other.

Is metric “dram_read_throughput” valid in Nsight Compute?

It is not. The naming format of that metric indicates it is a nvprof metric. The nvprof metric names can generally not be used directly in Nsight Compute. To find out if there is an "equivalent" metric in nsight compute for a given nvprof metric, use the nvprof transition guide, in particular the metric comparison table. By studying that table, you'll note that there is a Nsight compute metric that is equivalent to dram_read_throughput and it is named dram__bytes_read.sum.per_second For instructions on how to capture this metric in nsight compute, please refer to the blog I already mentioned here, or refer to the documentation here.

But I can not launch the above two software because of the lack of graphical interface. How could I use Nsight Compute in remote server?

If you have the CUDA toolkit installed on the remote server, you should be able to run Nsight Compute in CLI (command-line-interface) mode. That is described both in the documentation already linked, and the blog article already linked. Alternatively, you may be able to run the GUI in remote mode, as described here.

By the way, is it possible to profile metrics in Nsight Compute?

Yes, we have already covered that.

I won't be able to use this question/answer to debug remote connection details or any other follow-up questions about specific access cases or usage scenarios of Nsight tools. There are documentation and tutorials already available. If you have another specific question, please ask a new question. To locate resources for Nsight Compute and Nsight Systems, I suggest simply googling those names. Usually the first hits will be landing pages here and here which link to all of the above resources, plus additional resources such as video tutorials describing specific cases and advanced usage.

All of these tools are available on windows as well with similar user interfaces. Furthermore, these tools can/should be used for any GPU of compute capability 7.0 or higher.