Software Technology for Multi-core: Find parallel errors using MTRAT

Multicore architectures have become very popular. To take full advantage of such processor architectures, developers have to introduce parallelism in their software. Developing parallel software is hard and error prone. The most common parallel errors are data race and deadlock. When these errors occur, they are typically hard to locate. The Amino: Multi-threaded Runtime Analysis Tool (MTRAT) for Java can be used by developers to find these hard-to-find bugs.

Here I give two examples on how MTRAT finds data race and deadlock errors in parallel programs.

Data race occurs when two threads access the same memory location with no ordering constraints enforced between the accesses, where at least one of these accesses is a "write”. The Java sample below has potential data race. Class Value declares an integer field “x”, a synchronization method “add” to get x value and modify its value, a non-synchronization method “get” to return x value. We construct class Task by two Value instance and start two Task thread in main function. Run sample.DataRace with MTRAT, we can check potential data races in program runtime.

MTRAT detection result shows below. One data race is found by MTRAT, and it is reported with class fields, threads which access this field, and stack back trace information. This Read/Write data race happens on field x class sample/Value because two Task threads will visit instance v1 and v2, but add and get method in class Value are not protected by a same lock.

Run MTRAT in eclipse:

A deadlock is a situation in which two or more competing actions are waiting for the other to finish, so that neither ever does. Both situations are destructive and are common when building multi-threaded programs. The Java sample below has potential deadlock. In method harness2 of class Deadlock, two instances of StringBuffer L1 and L2 are created as parameters of class T3 constructor. In method run of class T3, thread will acquire two object locks sequentially, and then release them in opposite order. Since L1 and L2 are passed to t1 and t2 in opposite order, two threads will acquire two locks in opposite order, deadlock occurs. Run sample.Deadlock with MTRAT, we can check potential deadlocks in program runtime.