Saturday, January 9, 2016

How to Measure Cache Usage


Prime + Trigger + Probe


[1] demonstrates how to utilize the Prime + Probe technique to measure the cache activity, and extend it to the following Prime + Trigger + Probe measurement to support the setting of time-shared virtual machines.

  • The probing instance first allocates a continuous buffer B of b bytes
    • Here b should be large enough that a significant portion of the cache is filled by B.
  • Let s be the cache line size, in bytes.
  • Then the probing instance performs the following steps to generate each load sample
    1. Prime: Read B at s-byte offset in order to ensure it is cached. 
    2. Trigger: Busy-loop until the CPU's cycle counter jump by a large value
      • This means our VM was preempted by the Xen scheduler, hopefully in favor of the sender VM. 
    3. Probe: Measure the time it takes to again read B at s-byte offsets. 
When reading b/s memory locations in B, we use a pseudorandom order, and the pointer-chasing technique to prevent the CPU's hardware prefetcher from hiding the latencies. 

The time the of the final step's read is the load sample, measured in number of CPU cycles. These laod samples will be strongly correlated with use of the cache during the trigger step, since last usage will evict some portion of the buffer and thereby drive up the read time during the probe phase. 



Reference 


[1] Hey, You, Get Off of My Cloud: Exploring Information Leakage in Third-Party Compute Clouds, by Restenpart, T. et al., in CCS09

No comments:

Post a Comment