XRT: An Accelerator-Aware Runtime for Accelerated Chip Multiprocessors
summary
XRT is a runtime that runs on accelerated CMPs and claims up to 3.2× higher throughput-under-SLO versus an unoptimized runtime by distributing accelerator notifications to workers, avoiding dispatcher bottlenecks, and falling back to CPU when accelerators are overloaded. Evaluation on a 5th‑gen Intel Xeon (26 cores, two IAA and two DSA accelerators) shows XRT uses a monitoring set plus a notification-aware scheduler and a software fallback to prevent 55% drops in useful CPU work and ~13% context‑switch overheads, improving P99.9 latency behavior.
AI-written summary · about
chapters (15)
- 0:00XRT introduction
- 1:01Acceleration trends
- 2:47Integrating accelerators challenges
- 3:52XMP runtime patterns
- 5:15Dispatcher vs worker
- 6:40Overload risks
- 7:46Context switch overhead
- 8:27XRT architecture
- 9:53Software fallback
- 10:36Evaluation setup
- 11:19Comparison with alternatives
- 12:00Message processing results
- 13:05Limited acceleration workloads
- 13:46Accelerator bottleneck impact
- 14:06Conclusion
related talks
13mCohort: Software-Oriented Acceleration for You, Me, and Our Heterogeneous SoCs
25mOpen Source Software Stack for Heterogenous embedded devices
56mWrite Fast Code Like a Native
39mSoftware Defined Hardware on Heterogeneous SoCs Using OpenAMP
38mRemote Processor Framework in Zephyr RTOS
01h 21mAfter Reflection: The Runtime Story
01h 50mUnlocking Modern CPU Power - Next-Gen C++ Optimization Techniques
14mFreedom from memory interference