Research Program
Application Domains
New Software and Platforms
Bilateral Contracts and Grants with Industry
Partnerships and Cooperations
Bibliography
 PDF e-Pub

## Section: New Results

### Compilation and Optimization

Participants : Loïc Besnard, Caroline Collange, Byron Hawkins, Erven Rohou, Bahram Yarahmadi.

#### Optimization in the Presence of NVRAM

Participants : Erven Rohou, Bahram Yarahmadi.

A large and increasing number of Internet-of-Things devices are not equipped with batteries and harvest energy from their environment. Many of them cannot be physically accessed once they are deployed (embedded in civil engineering structures, sent in the atmosphere or deep in the oceans). When they run out of energy, they stop executing and wait until the energy level reaches a threshold. Programming such devices is challenging in terms of ensuring memory consistency and guaranteeing forward progress.

##### Checkpoint Placement based Worst-Case Energy Consumption

Previous work has proposed to insert checkpoints in the program so that execution can resume from well-defined locations. We propose to define these checkpoint locations based on worst-case energy consumption of code sections, with limited additional effort for programmers. As our method is based upon worst-case energy consumption, we can guarantee memory consistency and forward progress.

This work has been presented at the Compas 2019 conference.

Previous work has proposed to back-up the volatile states which are necessary for resuming the program execution after power failures. They either do it at compile time by placing checkpoints into the control flow of the program or at runtime by leveraging voltage monitoring facilities and interrupts, so that execution can resume from well-defined locations after power failures. We propose for the first time a dynamic checkpoint placement strategy which delays checkpoint placement and specialization to the runtime and takes decisions based on the past power failures and execution paths that are taken. We evaluate our work on a TI MSP430 device, with different types of benchmarks as well as different uninterrupted intervals, and we measure the execution time. We show that our work can outperform compiler-based state-of-the-art with memory footprint kept under the control.

This research is done within the context of the project IPL ZEP.

#### Dynamic Binary Optimization

Participant : Erven Rohou.

##### Guided just-in-time specialization

JavaScript's portability across a vast ecosystem of browsers makes it today a core building block of the web. Yet, building efficient systems in JavaScript is still challenging. Because this language is so dynamic, JavaScript programs provide little information that just-in-time compilers can use to carry out safe optimizations. Motivated by this observation, we propose to guide the JIT compiler in the task of code specialization. To this end, we have augmented [17] the language with an annotation that indicates which function call sites are likely to benefit from specialization. To support the automatic annotation of programs, we have introduced a novel static analysis that identifies profitable specialization points. We have implemented our ideas in JavaScriptCore, the built-in JavaScript engine for WebKit. The addition of guided specialization to this engine required us to change it in several non-trivial ways. Such changes let us observe speedups of up to 1.7$×$ on programs present in synthetic benchmarks.

##### Run-time parallelization and de-parallelization

Runtime compilation has opportunities to parallelize code which are generally not available using static parallelization approaches. However, the parallelized code can possibly slowdown the performance due to unforeseen parallel overheads such as synchronization and speculation support pertaining to the chosen parallelization strategy and the underlying parallel platform. Moreover, with the wide usage of heterogeneous architectures, such choice options become more pronounced. We consider [22] an adaptive form of the parallelization operation, for the first time. We propose a method for performing on-stack de-parallelization for a parallelized binary loop at runtime, thereby allowing for rapid loop replacement with a more optimized one. We consider a loop parallelization strategy and propose a corresponding de-parallelization method. The method relies on stopping the execution at safe points, gathering threads' states, producing a corresponding serial code, and continuing execution serially. The decision to de-parallelize or not is taken based on the anticipated speedup. To assess the extent of our approach, we have conducted an initial study on a small set of programs with various parallelization overheads. Results show up to 4$×$ performance improvement for a synchronization intense program on a 4-core Intel processor.

With the multicore trend, the need for automatic parallelization is more pronounced, especially for legacy and proprietary code where no source code is available and/or the code is already running and restarting is not an option. We engineer [21] a mechanism for transforming at runtime a frequent for-loop with no data dependencies in a binary program into a parallel loop, using on-stack replacement. With our mechanism, there is no need for source code, debugging information or restarting the program. Also, the mechanism needs no static instrumentation or information. The mechanism is implemented using the Padrone binary modification system and pthreads , where the remaining iterations of the loop are executed in parallel. The mechanism keeps the running program state by extracting the targeted loop into a separate function and copying the current stack frame into the corresponding frames of the created threads. Initial study is conducted on a set of kernels from the Polybench workload. Experimental results show from 2$×$ to 3.5$×$ speedup from sequential to parallelized code on four cores, which is similar to source code level parallelization.

This research was partially done within the context of the project PHC IMHOTEP.

#### Automatic and Parametrizable Memoization

Participants : Loïc Besnard, Erven Rohou.

Improving execution time and energy efficiency is needed for many applications and usually requires sophisticated code transformations and compiler optimizations. One of the optimization techniques is memoization, which saves the results of computations so that future computations with the same inputs can be avoided. We propose [16] a framework that automatically applies memoization techniques to C/C++ applications. The framework is based on automatic code transformations using a source-to-source compiler and on a memoization library. With the framework users can select functions to memoize as long as they obey to certain restrictions imposed by our current memoization library. We show the use of the framework and associated memoization technique and the impact on reducing the execution time and energy consumption of four representative benchmarks. The support library is available at https://gforge.inria.fr/projects/memoization (registered with APP under number IDDN.FR.001.250029.000.S.P.2018.000.10800).

#### Autotuning

Participants : Loïc Besnard, Erven Rohou.

The ANTAREX FET HPC project relies on a Domain Specific Language (DSL) based on Aspect Oriented Programming (AOP) concepts to allow applications to enforce extra functional properties such as energy-efficiency and performance and to optimize Quality of Service (QoS) in an adaptive way. The DSL approach allows the definition of energy-efficiency, performance, and adaptivity strategies as well as their enforcement at runtime through application autotuning and resource and power management. We present [20] an overview of the key outcome of the project, the ANTAREX DSL, and some of its capabilities through a number of examples, including how the DSL is applied in the context of the project use cases. We demonstrated [30] tools and techniques in two domains: computational drug discovery, and online vehicle navigation.

#### Loop splitting

The loop splitting technique takes advantage of long running loops to explore the impact of several optimization sequences at once, thus reducing the number of necessary runs. We rely on a variant of loop peeling which splits a loop into into several loops, with the same body, but a subset of the iteration space. New loops execute consecutive chunks of the original loop. We then apply different optimization sequences on each loop independently. Timers around each chunk observe the performance of each fragment. This technique may be generalized to combine compiler options and different implementations of a function called in a loop. It is useful when, for example, the profiling of the application shows that a function is critical in term of time of execution. In this case, the user must try to find the best implementation of their algorithm.

This research was partially done within the context of the ANTAREX FET HPC collaborative project, collaboration is currently ongoing with University of Porto, Portugal.

#### Hardware/Software JIT Compiler

Participant : Erven Rohou.

Single-ISA heterogeneous systems (such as ARM big.LITTLE) are an attractive solution for embedded platforms as they expose performance/energy trade-offs directly to the operating system. Recent works have demonstrated the ability to increase their efficiency by using VLIW cores, supported through Dynamic Binary Translation (DBT) to maintain the illusion of a single-ISA system. However, VLIW cores cannot rival with Out-of-Order (OoO) cores when it comes to performance, mainly because they do not use speculative execution. We study [27] how it is possible to use memory dependency speculation during the DBT process. Our approach enables fine-grained speculation optimizations thanks to a combination of hardware and software. Our results show that our approach leads to a geo-mean speed-up of 10 % at the price of a 7 % area overhead.

Our previous work on Hybrid-DBT was also presented at the RISC-V workshop in Zürich, Switzerland [38].

This work is a collaboration with the CAIRN team.

#### Scalable program tracing

Participants : Byron Hawkins, Erven Rohou.

The initial goal of scalable tracing is to record long executions at under 5$×$ overhead (ideally 2$×$), but it is equally important for analysis of the compressed trace to be efficient. This requires careful organization of the recorded data structures so that essential factors can be accessed without decompressing the trace or comprehensively iterating its paths. Precise context sensitivity is especially important for both optimization and security applications of trace-based program analysis, but scalability becomes challenging for frequently invoked functions that have a high degree of internal complexity. To avoid state space explosion in the context graph, such a function can be represented as a singleton while its complexity is preserved orthogonally. The current efforts focus mainly on developing an integration strategy to simplify program analysis over these two orthogonal dimensions of the trace.

#### Compiler optimization for quantum architectures

Participant : Caroline Collange.

In 2016, the first quantum processors have been made available to the general public. The possibility of programming an actual quantum device has elicited much enthusiasm [34]. Yet, such possibility also brought challenges. One challenge is the so called Qubit Allocation problem: the mapping of a virtual quantum circuit into an actual quantum architecture. There exist solutions to this problem; however, in our opinion, they fail to capitalize on decades of improvements on graph theory.

In collaboration with the Federal University of Minas Gerais, Brazil, we show how to model qubit allocation as the combination of Subgraph Isomorphism and Token Swapping [31]. This idea has been made possible by the publication of an approximative solution to the latter problem in 2016. We have compared our algorithm against five other qubit allocators, all independently designed in the last two years, including the winner of the IBM Challenge. When evaluated in “Tokyo”, a quantum architecture with 20 qubits, our technique outperforms these state-of-the-art approaches in terms of the quality of the solutions that it finds and the amount of memory that it uses, while showing practical runtime.