Department of Information Technology
Uppsala Architecture Research Team
uart_overview_flow.png

Fast Modeling

model.png Our modeling technologies allow us to rapidly explore different architectural parameters and analyze application behavior with very low-overhead. We can thereby take advantage of our unique measurement capabilities to rapidly and accurately predict behavior across a wide range of hardware and software configurations.

Statistical Cache Modeling

Statistical cache modeling is an extremely fast method for predicting an application's cache miss ratio from cheap (20% overhead) sampled memory reuse data. With these models we can instantly predict cache miss ratios for arbitrary sized caches.

Data cache miss ratio as a function of cache size and program phase.

Profile-based Contention Modeling

Leveraging the profile data collected from our Cache Pirate measurement tool, we can rapidly model cache contention effects, including all the details of the actual hardware and application.

Variability in execution speed due to different overlapping.

Power Modeling

Our low-overhead, performance-counter based power models allow us to understand and analyze program power behavior. By calibrating the models for commodity hardware, we can predict and improve power efficiency.

Measured and predicted power consumption for maximum and minimum frequencies based on data from maximum frequency execution.

High Performance Simulation

Our high-performance simulation frameworks allow us to understand and analyze full-system behavior of realistic (large) workloads. By exploiting hardware virtualization and performance sampling, our methods can reach execution rates comparable to native execution.

Execution times when running native compared to different simulation modes (projected).

gltracesim_overview.png

Our new graphics tracing and replay framework allows us to explore system-level effects on heterogeneous CPU+GPU memory systems. By efficiently generating GPU memory access traces for modern graphics applications, GLTraceSim replays them through high-level models and detailed simulators to explore effects in bandwidth, cache misses, scheduling and performance. GLTraceSim is built upon well-maintained, publicly-available tools.

Updated  2017-09-25 12:10:41 by Germán Ceballos.