Hi, Mikhail: These days i have read a lot of paper about dynamic optimization. I got the following ideas to profile the drlvm with PMU: 1. Prefetch Injection Based on Hardware Monitoring and Object Metadata: this paper is a good example about how to PMU to dynamic prefetch data with high cache miss rate. I want to implement this idea in DRLVM. 2. The idea from Chilimbi. He got the trace by instrumented the code, which causes high overhead. I want to improve his implementation. Firstly use the hardware performance counter to get the data reference trace, and abstracts the trace. Secondly use SEQUITUR algorithm to simplify the trace and represent the trace with DAG from which we can easily get the hot data stream (follow Chilimbi`s way). This can instruct how to prefetch the data with high miss rate.
As for the performance monitor tool, i will think about it later. Please point out the defect or mistake. Thanks.
