Download presentation
Presentation is loading. Please wait.
Published byCrystal Emma Campbell Modified over 6 years ago
1
Dagstuhl Seminar on Dark Silicon: From Embedded to HPC Feb 3, 2016
HPC Group Dagstuhl Seminar on Dark Silicon: From Embedded to HPC Feb 3, 2016
2
What is Dark Silicon in HPC?
Shortage of power, strict constraints Overprovisioning of hardware resources with respect to power Chip-level: we already have dark silicon Job-level, cluster-level: workload dependent, dim silicon
3
Main challenges arising from DS?
Focus in HPC is performance, we care about FLOPS, time to solution, accuracy of science. Different than the embedded community Stay in a power band, fluctuations of a few megawatts may arise, power stability Performance variability, reproducibility System throughput and power utilization: how do we manage resources?
4
Challenges… Performance Variability: Embrace it and build runtime systems to address this At the cost of performance? Use it to our advantage: need more control? Minimize in hardware? How do we tolerate variability? Current approaches for programming the machines and performance tuning will no longer work. We need to redefine the metrics on how we measure things. We assume homogeneity at the moment.
5
Fundamental Techniques
Monitoring at large scale (TUM, LRZ, LLNL, ETH, Dresden, JSC …) Granularity, what to gather, how to use it Resource Management, RMAP/Flux (LLNL, TUM, ETH, UniBo…) Runtime Systems GEO PM (Intel), Conductor (LLNL), SDSC, PowSched (UOregon) Tracing, Visualization (LLNL, JSC, Dresden, Oregon,SDSC, …) Deeper connection with workflows and applications. Also, refactoring of applications (work stealing, oversubscription, auto tuning)
6
Opportunities and Impacting the Embedded community?
Measurement and Analysis Tools Learning from the operating systems community about hardware management Real-time boundaries, can we introduce them in HPC, have a QOS definition? Programming models tailored to code (right techniques for programming) Portability, development cycle Hardware-software co-design Redesign the stack: hardware, job, cluster, site
7
Going Forward: Big Data and HPC
HPC Applications are handling more data, which creates a new set of challenges… Data-intensive computing, data-aware computing Another Dagstuhl?
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.