A page for coordinating project groups and ideas.
Mattan will add ideas if the list isn’t long and diverse enough :-)
Ideally, the project will be something that aligns with what you are already doing (in research or at work).
I’d like three per group but do make exceptions for good reasons.
Project Groups
| Group Members | Project Topic | |
|---|---|---|
| 1 | Gokhan, Mike, Hyungman | |
| 2 | Tianhao, Jee Ho, Michael | |
| 3 | Jinsuk, Minsoo, Jungrae | |
| 4 | Ben, Dan, Joshua | “Jumpstarting a branch predictor” |
| 5 | Ali Unwala, Shane Mahon, Daniel Arulraj | |
| 6 | Haishan, Derong, Song | |
| 7 | ||
| 8 |
Ideas
Just list ideas you want to work on and contact info in hopes of quickly forming groups. Mattan will add some ideas at some point (and if pinged repeatedly).
Integrator system: Electric vehicles providing renewable energy following and regulation services
Contact: Mike Legatt (mlegatt@utexas.edu)
Description: Electric vehicles have the potential to not only improve local air quality (less than 50% the CO2 emissions driving off coal/natural gas power plant energy than petroleum), but they can also be used to improve the reliability of the bulk power system. They can be used to offset intermittent renewables, and even to offer ancillary services, increasing and decreasing vehicle charging at times to support the grid (< 5 second responsiveness). In order to do this, a high-reliability/speed linkage is necessary between an ISO or utility, EVSE (vehicle charger) and EV, with GPS-resolution timestamp data. To this end, a high-speed bidirectional network connected EVSE is proposed, connected to a synchrophasor, providing reliability measurements back to the ISO (e.g., voltage, frequency, phase angle against a GPS times tap), as well as real-time vehicle state of charge information and billing-grade telemetry. These high-speed services require high availability and responsiveness, but also need to incorporate end-user interfacing via mobile device, and an aggregation system linking location, vehicle, charging station, driver preferences and long-term scheduling together.
Device control from programmable GPUs
Contact: Michael Lee <mzlee@cs.utexas.edu>
Description: While programmable GPUs are great for heavy parallel processing, the data needs to come from some source. This makes at least some of the compute time bound by the CPU and its ability to process network packets, access the file system, or stream data from input devices like cameras. However, some of these sources are inherently data parallel and processing efficiency could benefit from not having to go through the CPU and CPU memory just to be copied again.
One thought along these lines is to try to understand what kinds of architectural changes could allow more efficient processing. For example, allowing a kernel to spawn a sub-kernel in response allows the programmer to avoid a GPU to CPU communication. Are there other similar, supportable constructs that could help improve performance for a different type of workload?
Jumpstarting a Branch Predictor: minimizing cold start branch mispredictions after thread migration
Contact: Ben Lin (bencplin@utexas.edu)
Description: Asymmetric Chip Multiprocessors(ACMP) have both large, out-of-order cores (designed for single-threaded performance), and smaller, in-order cores (designed for energy/power efficiency). The idea is to accelerate multi-threaded applications on an ACMP by executing the program bottlenecks (i.e. serial segments, critical sections, lagging threads to a barrier, etc.) on the large core, while executing non-bottlenecks on energy efficient cores. However, if a bottleneck is short, the speedup from executing on a large core can be negated by cold start effects in the cache and branch predictor. Existing work has already addressed the issue of cold misses in the cache via Data Marshalling. However, the issue of cold start branch mispredictions has not been studied as extensively. The only previous solution we’ve come across is to ship the global history register with the bottleneck to the large core, which does reduce the cold start mispredict penalty, but not by a lot. We believe we can come up with a more effective solution and plan to evaluate both its effectiveness and feasibility.
