Activities in the Variability Expedition,  a multi-institutional collaborative project, focused on developing hardware and software technologies, algorithms, and foundational theories necessary for a new class of computing machines that operate efficiently and reliably in the presence of unprecedented variations in their hardware characteristics arising from manufacturing, operating conditions, application workloads, and aging. During the project, the UCLA researchers made the following contributions towards the project’s objectives, and developed tools, methods, and software artifacts that have been adopted by companies and used by other researchers.

1. Computing platforms capable of sensing variations

For software to adapt to variations in the characteristics of underlying hardware, a computer must have circuits that can sense those variations. Project researchers developed a variety of delay variation sensing schemes and tested them in simulation as well as in silicon prototypes (Figure 1)). First, they developed ‘Design Dependent Ring Oscillators’  and automated algorithms to synthesize them as an inexpensive circuit structure that mimics performance of the actual chip. Second, they developed ‘SlackProbe’ scheme for generating minimum number of in-situ delay monitors which can minimize monitoring overhead by over 10x as observed on commercial microprocessor designs.

2. Variation-aware Memory Management

Memory circuits and subsystems where software stores intermediate results of computation are particularly susceptible to variations which cause degradation in memory reliability and energy efficient. Project researchers developed a suite of techniques (Figure 2) to opportunistically exploit memory variability for energy savings and cope with memory errors when they inevitably occur. First, they developed ‘ViPZonE’ and ‘DPCS’ which save energy in off-chip main memory and on-chip cache memories, respectively. Second, they developed ‘X-Mem’ a new extensible memory characterization tool and ‘MEMRES’ a fast large scale memory fault simulator. Finally, they demonstrated that correctable memory errors can have a significant impact on performance of datacenter computing systems.

3. Theory-informed coding algorithms to exploit variations in memory and storage

Project researchers developed a broad repertoire of novel error correction schemes, including tensor product codes and non-binary coding, to exploit variations in memory and storage. These schemes are based on advanced mathematical techniques introduced by them using tools from abstract algebra, combinatorics, and graphical methods. In contrast to long standing, conventional information theoretic methods that assume symmetric error correction, these new schemes explicitly take advantage of the underlying hardware variability to design provably better error corrections schemes. They offer impressive lifetime improvements (Figure 3)  in modern memory devices, as well as theoretical guarantees that can be applied in other applications.

4. Variability-aware duty-cycling for ultra-low-power embedded sensors

UCLA researchers experimentally identified that manufacturing variations in turn result in more than an order of magnitude variation in sleep-mode power consumption of embedded computing devices such as wireless sensors. Based on this insight they developed open-source software packa  VaRTOS (a real-time operating system kernel that provides abstractions that explicitly handle idle and active power variations), ViRUS (a run-time for supporting variability aware applications based on the concept of virtual function replacement), and VarEmu (a simulation framework for the evaluation of variability-aware software techniques).

5. Context variation aware software stack for smartphones

Researchers at UCLA developed CAreDroid, an open-source software framework for enabling   applications under Android mobile operating system to efficiently adapt to variations in platform resources and context. CAreDroid (Figure 5),  implemented as part of the Android runtime system, allows developers to focus on the application logic by decoupling application logic from run-time tasks such as monitoring of user and platform state, and adaptation of process, battery, and other resources. It enables applications that are more energy efficient, robust, and easier to develop.

6. Los Angeles Computing Circle

As part of the project the team members at UCLA created The Los Angeles Computing Circle, (http://lacc.ee.ucla.edu) an outreach program organized, supervised, and mentored by faculty members and graduate student volunteers with an objective of engaging and mentoring younger students for careers in computing and engineering. LACC, organized each summer as a sequence of self-contained modules,  provides Grade 9-12 students an opportunity to learn advanced concepts in computing via lectures and hands-on design and programming laboratories. The program complements and go beyond the more abstract exposure to computing that the traditional high school computer science curriculum offers by way of AP Computer Science and related courses. Alumni of LACC have gone on to pursue advanced STEM degrees at top universities.

Last Modified: 02/12/2018
Modified by: Mani B Srivastava