Review of Reducing Memory Interference with Memory Channel Partitioning

Review of Reducing Memory Interference in Multicore Systems via Application-Aware Memory Channel Partitioning

 

Paper Summary

The paper focus on the problem that in multicore systems if the interference between different applications are not controlled, it will severely influence the system performance. While existing work aim to mitigate the problem by changing the scheduling policy and prioritizing the memory requests in a way that improve performance, the authors propose two new methods: a hardware method partitioining the channels to different applications, and a software method by priortizing very little requests.

In the first hardware method, application-aware memory channel partitioning(MCP), they seperate the requests into memory intensive and memory non-intensive groups, and the intensive group is divided into high row-locality and low row-locality in a finer level. Number of channels assigned to applications in different group are calculated differently considering the group charicaristics. 

In the second software method, integrated memory partitioning and scheduling(IMPS), the software can previously prioritize the very little requests, because they are very small and can finish very soon, avoiding further interfering with other applications.

With the experiments the authors show that, 7.4% and 11.1% over baseline, 1% and 5% over previous best scheduler throughput improvement are achieved by MCP and IMPS individually, while they require less hardware complexity. MCP and IMPS can be used toghether to get higher performance, and they can even be combined with other previous work to get further improvement.

Strengths

1. The methods are in a different perspective, from partitioning the channels and in software, which is appealing.

2. The methods are complementary to existing work and can further improve existing work.

3. The experiments are solid and adequate, and many sensitivity work are done.

4. For some strange situations where MCP and IMPS are defeated by previous work, explaination are given.

5. Comparing shows that the strengths are simple complexity and throughput, but sacrificing fairness.

Weakness

1. Fairness

2. says high-intensity workload will be the trend but not giving any evidence.

3. when page may be present in a channel not expected, may happen at 12% probability, the author does not concern it because it is low, but as page fault may inquire many cycles of stalling, it's still an unavoidable loss of performance.

4. still not solving the possibility that in condition of proportional allocation channel, when some group outweight, the small group have 0 channels. Even though the probability is low, it should be solved not to induce errors.

5. calculating the number of channels in a group may be wrong.......

Ideas

As the author mentioned in the work, MCP and IMPS are based on FR-FCFS but not some modern scheduling policy, so combining them to FR-FCFS can get much more improvement than ATLAS and TCM. The author mention that scheduling are working with heterogeous requests, but the paper try to group them and prioritize requests in the group. I think some in-group prioritization(almost equal to scheduling) can be done and the MCP\IMPS better utilizied by ATLAS and TCM.