- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I am trying to measure the L2 -> L1 Fill latency.
I'm using PMC 0x48 umask=0x01. This increments each cycle the number of Line Fill Buffers currently allocated, right?
If I also tabulate the number of allocations in the L1D (PMC 51 umask=0x01):
<# cycles a fill takes from L2 -> L1> =(# active LFB per instruction) / (# L1D allocations made per instruction)
Is this correct?
perfwise
Link Copied
1 Reply
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hello perfwise,
PMC 0x48, umask 0x1 (L1D_PEND_MISS.PENDING) incrementsthe number of outstanding L1D misses every cycle.
The counting probably starts as soon as an L1D miss occurs and before a buffer is allocated for the miss.
The L1Dmiss might require a trip to L2 or L3 or memory. I suspect that it counts bothcacheable and uncacheable memory accesses.
L1D_PEND_MISS contains an unknown mix of types of misses.
So you can't use it to compute an L2->L1 miss latency.
Pat
Reply
Topic Options
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page