Global_op_st.ratio metric

Hello,
I see two metrics
smsp__sass_average_data_bytes_per_sector_mem_global_op_st.pct [byte/%]
and
smsp__sass_average_data_bytes_per_sector_mem_global_op_st.ratio [byte/byte]
however, in the document page I only see smsp__sass_average_data_bytes_per_sector_mem_global_op_st.pct is the equivalent of gst_efficiency in nvprof. So, the question is what is that *.ratio metric?

They are the same metric, but once expresses the value as a percentage, and one as a ratio.

1 Like

OK but the ratio is not clear. For example, I see
.pct = 31.2
.ratio = 10
So, the former means gst_efficiency is 31.2% but I don’t understand the latter. ratio of st to what?

It can be derived from the metric name. Here, it is he ratio of (average) data bytes per sector (for store operations accessing the global memory space).

1 Like

One more question.
I see that sometimes the pct value is larger than 100. For example:
.pct = 101.4
.ratio = 32.46
So, I wonder how the percentage above 100 is interpreted. Is that normal?

Small deviations from the logically correct range are an expected measurement error if a metric is composed of multiple counters, and those are collected across different replay passes. Having clock and cache control enabled helps to reduce the error, but it can still occur due to non-determinism in the kernel or HW. You can try to further reduce the measurement error by collecting the metric by itself, but there is no guarantee if it’s not collectable in a single replay pass.

You can refer to https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#range-and-precision for further details.

1 Like