../vtmm-experiments

vTMM Experiments

Environment setup

Section Structure

titlecolumns
1experimental setup
2experimental approach3/4
3page tracking efficiency2+1/2
4page migration efficiency1+3/4
5ablation study (page classification)1+1/4
6fixed/dynamic hot set1
7sensitivity to dram size1
8thp support1+3/4
9multi-vm co-running2+1/2

In summary, the structure is as expected. It contains micro and macro benchmarks. Macro benchmarks demonstrate overall system performance and micro benchmarks shows performance of every design decision. The demonstrated designs covers collection, identification, migration and rebalancing. In particular, it contains more details on collection and rebalancing.

Tracking (4.3)

To demonstrate the accuracy of page table scanning, the authors compared EPT and GPT scanning. The results strongly favours vTMM, because page table scanner scans mapped page, even when guest pages are not used, they are still mapped in the host.

This Benchmark will favour ours even more, because we are a sample based approach, the scanned memory will only contains captured pages.

To demonstrate the the accuracy of multi-level queue, the author demonstrated the distribution of page access count (figure 3). Although the hot region has more accesses when MLQ is enabled, the difference is miniscule.

figure 3: accuracy of the multi-level queue

Problems:

Classification

Migration (4.4)

table 4: read slowdown caused by migration table 5: write slowdown caused by migration table 6: VMTraps caused by page migration

Problem:

Improvement breakdown

performance breakdown