Google Summer of Code
From gem5
All the ideas listed here will require some familiarity with Python and a good grasp of advanced C++ concepts.
- Build a direct execution CPU model based on the Linux Kernel Virtual Machine
- Parallelize M5
- Use the Wisconsin Wind Tunnel as a guide
- This actually isn't as bad as it sounds as all objects schedule their own events and there are limited ways they can interact with other objects in the system.
- Crossbar network
- Mesh network
- Directory Protocol
- Real F/S In-order core
- Kevin has one that works to some degree in SE, it doesn't have functional units yet, but it does have variable latency stages and such
- Korey has one he did at MIPS, I don't know about it's features, but it's SE only as well
- SMT
- Fix O3 bugs/ Fix ROB Units
- It's a huge pile of code to understand before you get anywhere if they get that far
- Are there any other benchmarks we want?
- That they could possible make work?
- Get and intergate one of the various other DRAM models people have
- Implement a more complete version of the Intel NIC we simulate
- TSO, enforce the number of DMA engines, etc.
- Sampling/checkpointing/restarting
- Testing, fixing, etc... Not exactly exciting work
- Write a PLI interface to connect Verilog CPUs to the memory system.
- Sampling/fast-forwarding techniques (making sure ours works, maybe adding in some new ones)
- Different cache models (different replacement policies, etc.; would allow them to do some research into it, maybe get some work done for Lisa)
- Different prefetcher models (expand on what Ron has, also can do some research into it)
- Flash memory device model? (seems popular nowadays)
