Erik Kruus

Static data race detection for concurrent programs with asynchronous calls

A large number of industrial concurrent programs are being designed based on a model which combines threads with event-based communication. These programs consist of several threads which perform computation by dispatching tasks to other threads via asynchronous function calls. These asynchronous function calls are implemented using function objects, which are essentially wrappers containing a pointer to the function that should be executed on a particular thread with the corresponding arguments. In many cases, the arguments, in turn, contain function objects which serve as callbacks. Verifying such programs which involves reasoning about complex concurrency constructs comprising function pointers and callback functions is extremely tricky especially in the presence of recursion. In this paper, we present a fast and accurate static data race detection technique for multi-threaded C programs with asynchronous function calls and demonstrate its application to real-life software.

MALT: distributed data-parallelism for existing ML applications

Machine learning methods, such as SVM and neural networks, often improve their accuracy by using models with more parameters trained on large numbers of examples. Building such models on a single machine is often impractical because of the large amount of computation required. We introduce MALT, a machine learning library that integrates with existing machine learning software and provides data parallel machine learning. MALT provides abstractions for fine-grained in-memory updates using one-sided RDMA, limiting data movement costs during incremental model updates. MALT allows machine learning developers to specify the dataflow and apply communication and representation optimizations. Through its general-purpose API, MALT can be used to provide data-parallelism to existing ML applications written in C++ and Lua and based on SVM, matrix factorization and neural networks. In our results, we show MALT provides fault tolerance, network efficiency and speedup to these applications.

EEffSim: A discrete event simulator for energy efficiency in large-scale storage systems

The complexity of distributed storage systems makes it difficult to evaluate fundamentally new data placement policies for energy efficiency in real-world systems. Simulation studies allow us to quickly prototype and test energy-efficient management policies. Unfortunately, there is a lack of tools in the public domain that support such studies. We introduce EEffSim, a highly configurable energy simulator for general-purpose multi-server storage systems. EEffSim supports data migration, locking protocols, write offloading, and opportunistic spin-down. It also models the energy consumption of heterogeneous storage devices including SSDs.