Performance test

[1a1a11a]

As we aggressively refactor the codebase, we need performance tests to ensure that new changes do not negatively affect performance, both on the CPU and DRAM. This needs to be integrated with the existing CI that uses dedicated server.

[1a1a11a]

@haochengxia do you have a rough idea of throughput running the twitter trace when we use

1. cachesim LRU
2. cachesim pluginLRU (C++)
3. Python LRU
4. Python PluginLRU?
   No need to run extensive experiments, just to get a rough idea

[haochengxia]

I plan to add a throughput benchmark script and a simple memory profiler driven by "valgrind massif". I think we'll get a rough evaluation soon.

Previously, I only tested some small traces from TencentBlock, which may not be very representative.

[1a1a11a]

Sounds good! We can also delegate this task to a new student, as it is relatively simple.

[haochengxia]

Sounds good! We can also delegate this task to a new student, as it is relatively simple.

Sure!

FYI: Python bindings have two methods to walk the trace, process_trace, which can specify the range for simulation, and for req in trace/reader, which can get request-level info. process_trace shows almost the same time costs as most of the operation happens on the C/C++ side, e.g., cachesim LRU 0.1s, process_trace 0.1s, walk through in request-level 0.4s.

[1a1a11a]

Is the 0.4s per request or full trace?

[haochengxia]

It is for full trace 😄

[haochengxia]

Post benchmark script plots from libCacheSim-python