Optimizing Threads Schedule Alignments to Expose the Interference Bug Pattern

Abstract

Abstract. Managing and controlling interference conditions in multi-threaded programs has been an issue of worry for application developers for a long time. Typically, when write events from two concurrent threads to the same shared variable are not properly protected, an occurrence of the interference bug pattern could be exposed. We propose a mathematical formulation and its resolution to maximize the possibility of exposing occurrences of the interference bug pattern. We formulate and solve the issue as an optimization problem that gives us (1) the optimal position to inject a delay in the execution flow of a thread and (2) the optimal duration for this delay to align at least two different write events in a multi-threaded program. To run the injected threads and calculate the thread execution times for validating the results, we use a virtual platform modelling a perfectly parallel system. All the effects due to the operating system's scheduler or the latencies of hardware components are reduced to zero, exposing only the interactions between threads. To the best of our knowledge, no previous work has formalized the alignment of memory access events to expose occurrences of the interference bug pattern. We use three different algorithms (random, stochastic hill climbing, and simulated annealing) to solve the optimization problem and compare their performance. We carry out experiments on four small synthetic programs and three real-world applications with varying numbers of threads and read/write executions. Our results show that the possibility of exposing interference bug pattern can be significantly enhanced, and that metaheuristics (hill climbing and simulated annealing) provide much better results than a random algorithm.

Using Contracts to Guide the Search-Based Verification of Concurrent Programs

“…Bhattacharya et al [2] use hill climbing and simulated annealing to identify potential race conditions through dynamic testing. A simulator is used to execute the program as this allows control of the thread interleaving by injecting timing delays while removing other sources of non-determinism.…”

ACRE: An Automated Aspect Creator for Testing C++ Applications

“…We thus create a DSL, detailed below, to describe the tests. The DSL does not [34] change the compilation or behavior of the program under test. We then create ACRE, a program that reads the DSL statements and generates the corresponding aspects.…”

“…However, it requires a complete enumeration of all possible schedules. Bhattacharya et al [34] proposed a mathematical approach to help these tools find interference bugs. Their approach relies on the exact times when threads write in a variable and we show in Subsection VI-C that ACRE can provide these times.…”

“…Finally, to answer RQ3, we use generated aspects to get access times for both reads and writes to a shared variable in NOMAD. These times are then provided to another approach, [34], to help finding possible interference bugs.…”

“…The corresponding aspect, shown in Listing 9, gives us all the access times for the attribute Mesh index, which is used in the MPI version of NOMAD. Once these times are recorded, we provide them as an input to the approach proposed by Bhattacharya et al [34], which gives us the optimal place to inject a delay (and its duration) in the source code so that we maximize the possibility of having an interference bug pattern. After injecting the given delay in the NOMAD source code, we found that indeed an interference bug may occurs for the attribute Mesh index.…”