Hi
all,
Phil
and I have been working on the PRD narrative for the Understanding section. I've attached our draft version below. The approach we took was the following:
1.
Copy in all the material from the corresponding PRD slide.
2.
First paragraph gets lightly edited into an intro.
3.
Top-level bullets become bolded topics.
4.
Sub-bullets become bullets to those.
5.
Added some narrative below that provides a little additional material/detail on those topics.
Lucy,
all: Does this look about right as an example?
Thanks!
Rob and Phil
---
5.5.6 Promising Research Directions
Participants concluded that better understanding of
SSIO systems leads to more productive applications, enhanced adaptation of running services, and more effective design, development, and procurement of tomorrow’s applications, systems, and system software. The following priorities emerged from the discussion.
Enabling real-time and post hoc
analysis through instrumentation, capture, and retention of monitoring data.
-
Scalable and minimally intrusive methods of data gathering on application-
and system-side, keeping pace with technological advancements in hardware and software
-
Streaming delivery of data for real-time analysis and decision-making
Scalable and minimally intrusive data gathering methods
enable the investigation of full-scale applications and systems rather than just experimental examples. This ensures that instrumentation data is representative of real-world behavior, and also enables real time feedback to users. For such instrumentation
to remain relevant, it must incorporate emerging technologies as they are deployed. This includes new hardware such as NVRAM, new programming models such as machine learning frameworks, and new architectures such as object storage systems. Furthermore, streaming
delivery of data will make the it applicable not just for post-hoc analysis, but for real time analysis and real time control feedback.
Predicting behavior through workload,
software stack, and architectural modeling.
-
Analytical, ML, and reduced models for rapid decision-making
-
Validated models at multiple fidelities for design space exploration
-
System models to aid in data integration
Modeling of relevant systems, services and workloads
complement data gathering activities. A variety of validated, predictive models should be considered depending on the use case for the model. This calls for further research in methods including analytical, machine learning, and reduced models to cover the
spectrum of high fidelity and rapid decision-making use cases. Additionally, integration of SSIO models with models capturing other aspects of the system (e.g., network and communication workload, scheduler) must be accounted for when tackling questions of
importance to facilities, both in terms of day-to-day operations and future procurement decisions.
Integrating data sources and identifying
correlations to improve our understanding.
Transformation of monitoring data into knowledge is
a nontrivial task that naturally involves fusion of data from multiple sources. Important real-time (e.g., autonomics) and post hoc (e.g., performance tuning) use cases exist with distinct needs. Machine learning is a promising approach to knowledge generation,
but for these approaches to be effective, the aforementioned challenges in instrumentation and capture must be addressed in order to gather the requisite data on which learning can occur.