Dram transactions (bytes) question!

Developer Tools Nsight Compute
1

I have a question about the DRAM sectors metric, specifically dram__sectors.sum. My understanding is that this value should represent the number of L2 cache misses.

However, when I try to calculate it by multiplying the total number of L2 sectors (lts__t_sectors.sum) by the L2 miss rate (1 - lts__t_sector_hit_rate.pct), the result doesn’t match the value reported by dram__sectors.sum.

Here’s an example from a GEMM application:
lts__t_sector_hit_rate.pct (%) 99.37
lts__t_sectors.sum (sector) 27,430,992
dram__sectors.sum (sector) 98,312

Shouldn’t dram__sectors.sum be equal to lts__t_sectors.sum * (1 - lts__t_sector_hit_rate.pct/100)?
Why is there a discrepancy between these values? Am I misunderstanding how these metrics are defined?

Any clarification would be appreciated!

2

Edited after realising I hadn’t read the question properly.

I think dram__sectors.sum measures only traffic read from DRAM, on the input side to the L2 cache.

3

dram__sectors.sum will include all read and write sectors at the memory controller.

lts__t_sectors.sum includes tag sector lookups which can hit or miss. This is focused on L2 data RAM accesses.
lts__t_sectors_lookup_miss.sum is not equivalent to dram__sectors.

Here are some considerations on the differences:

  • Requests to L2 that miss in L2 can result in both fill sectors and evict write-back sectors if selected line is dirty.
  • Writes (hit or miss) may generate write-through sectors to dram. This is not predictable as L2 is point of coherence a write-through is not always required for device memory.
  • Compressed reads/writes can result in additional dram traffic (on hit or miss).
4

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.