Section: New Results

Results on Heterogeneous and dynamic software architectures

We have selected three main contributions : two are in the field of runtime management, while the third one is in the field of non-functionnal software testing.

Precise and efficient resource management using models@runtime

Contribution.

We have developed an efficient monitoring framework to quickly spot an abnormal resource consumption within a complex application. In these papers [25], we have proposed an optimistic adaptive monitoring system to determine the faulty components of an application. Suspected components are finely analyzed by the monitoring system, but only when required. Unsuspected components are left untouched and execute normally.

Originality.

Current solutions that perform permanent and extensive monitoring to detect anomalies induce high overhead on the system, and can, by themselves, make the system unstable. Our system performs localized just-in-time monitoring that decreases the accumulated overhead of the monitoring system. Through our evaluation, we show that our technique correctly detects faulty components, while reducing overhead by  92.98 on average%.

Impact.

Beyond the scientific originality of this work, the main impacts of this novel approach approach to monitor software component performance has been to (i) reinforce DIVERSE's visibility in the academic and industrial communities on software components and (ii) to create several research tracks that are currently explored in different projects of the team (HEADS and B-com PhD thesis). This work has been integrated within the Kevoree platform.

Dynamic web application using models@runtime

Contribution.

We have developed a component-based platform supporting the development of dynamically adaptable single Web page applications. An important part of this contribution lies in the possibility to dynamically move code from the server to the client side allowing a great flexibility in the performance management. This contribution [56] is based on a models@runtime approach and has been implemented in our open source KevoreeJS platform.

Originality.

Current solutions to create single Web page application are limited to a static code repartition between clients and server, thus limiting the flexibility at runtime.

Impact.

Beyond the scientific originality of this work, the main impacts of this novel approach to monitor software component performance has been to (i) reinforce DIVERSE's visibility in the open-source community, (ii) to start several research tracks that are currently explored in different projects of the team (HEADS, STAMP, GRevis). This platforms is modular, one of the component has a monthly download count greater than 100k (https://www.npmjs.com/package/npmi)).

Testing non-functional behavior of compiler and code generator

Contribution.

We have developed NOTICE [36], [35], a component-based framework for non-functional testing of compilers through the monitoring of generated code in a controlled sand-boxing environment. In this work, we have proposed an automatic way of testing non-functional properties of compilers, while optimizing the generated application with respect to a set of specific non-functional properties (CPU, memory usage, energy consumption, etc.).

Originality.

Compiler users generally apply different optimizations to generate efficient code with respect to specific non-functional properties such as energy consumption, execution time, etc. However, due to the huge number of optimizations provided by modern compilers, finding the best optimization sequence for a specific objective and a given program is more and more challenging.

Impact.

Beyond the scientific originality of this work, the main impact of this novel approach is to enable the auto-tuning of compilers according to user requirements and to construct optimizations that yield to performance results that are better than standard optimization levels.

Automatic Microbenchmark Generation to Prevent Dead Code Elimination and Constant Folding

Contribution.

Microbenchmarking consists of evaluating, in isolation, the performance of small code segments that play a critical role in large applications. The accuracy of a microbenchmark depends on two critical tasks: wrap the code segment into a payload that faithfully recreates the execution conditions that occur in the large application; build a scaffold that runs the payload a large number of times to get a statistical estimate of the execution time. While recent frameworks such as the Java Microbenchmark Harness (JMH) take care of the scaffold challenge, developers have very limited support to build a correct payload. This year, we focus on the automatic generation of pay-loads, starting from a code segment selected in a large application [54]. In particular, we aim at preventing two of the most common mistakes made in microbenchmarks: dead code elimination and constant folding. Since a microbench-mark is such a small program, if not designed carefully, it will be over-optimized by the JIT and result in distorted time measures. Our technique hence automatically extracts the segment into a compilable payload and generates additional code to prevent the risks of over-optimization. The whole approach is embedded in a tool called AutoJMH, which generates payloads for JMH scaffolds. We validate the capabilities AutoJMH, showing that the tool is able to process a large percentage of segments in real programs. We also show that AutoJMH can match the quality of payloads handwritten by performance experts and outperform those written by professional Java developers without experience un microbenchmarking.

Collaborations

This year, we had a close and fruitful collaboration with the industrial partners that are involved in the HEADS and Occiware projects, in particular an active interaction with the Tellu company in Norway in the Heads context [49]. Tellu relies on Kevoree and KevoreeJS to build their health management systems. They will be also a active member the new Stamp project led by DIVERSE. We can cite also an active collaboration with Orange Labs through Kevin Corre's joint PhD thesis. Another joint industrial (CIFRE) PhD started in September 2016, and we are also partner in a new starting FUI project. Finally, DIVERSE collaborates with the B-COM IRT (https://b-com.com/en), as one permanent member has a researcher position of one day per week at B-COM and a new joint PhD started in September [52].

At the academic level we collaborate actively with the Spiral team at Inria Lille (several joint projects), the Tacoma team (with two co-advised PhD students), the Myriad team (1 co-advised PhD student) and we have started two collaborations with the ASAP team.