Transactional memory

Chip Multithreading (CMT) processors promise to deliver higher performance by running more than one stream of instructions in parallel. To exploit CMT's capabilities, programmers have to parallelize their applications, which is not a trivial task. Transactional Memory (TM) is one of parallel programming models that aims at simplifying synchronization by raising the level of abstraction between semantic atomicity and the means by which that atomicity is achieved. TM is a promising programming model but there are still important challenges that must be addressed to make it more practical and efficient in mainstream parallel programming. The first challenge addressed in this dissertation is that of making the evaluation of TM proposals more solid with realistic TM benchmarks and being able to run the same benchmarks on different STM systems. We first introduce a benchmark suite, RMS-TM, a comprehensive benchmark suite to evaluate HTMs and STMs. RMS-TM consists of seven applications from the Recognition, Mining and Synthesis (RMS) domain that are representative of future workloads. RMS-TM features current TM research issues such as nesting and I/O inside transactions, while also providing various TM characteristics. Most STM systems are implemented as user-level libraries: the programmer is expected to manually instrument not only transaction boundaries, but also individual loads and stores within transactions. This library-based approach is increasingly tedious and error prone and also makes it difficult to make reliable performance comparisons. To enable an "apples-to-apples" performance comparison, we then develop a software layer that allows researchers to test the same applications with interchangeable STM back ends. The second challenge addressed is that of enhancing performance and scalability of TM applications running on aggressive multi-core/multi-threaded processors. Performance and scalability of current TM designs, in particular STM desings, do not always meet the programmer's expectation, especially at scale. To overcome this limitation, we propose a new STM design, STM2, based on an assisted execution model in which time-consuming TM operations are offloaded to auxiliary threads while application threads optimistically perform computation. Surprisingly, our results show that STM2 provides, on average, speedups between 1.8x and 5.2x over state-of-the-art STM systems. On the other hand, we notice that assisted-execution systems may show low processor utilization. To alleviate this problem and to increase the efficiency of STM2, we enriched STM2 with a runtime mechanism that automatically and adaptively detects application and auxiliary threads' computing demands and dynamically partition hardware resources between the pair through the hardware thread prioritization mechanism implemented in POWER machines. The third challenge is to define a notion of what it means for a TM program to be correctly synchronized. The current definition of transactional data race requires all transactions to be totally ordered "as if'' serialized by a global lock, which limits the scalability of TM designs. To remove this constraint, we first propose to relax the current definition of transactional data race to allow a higher level of concurrency. Based on this definition we propose the first practical race detection algorithm for C/C++ applications (TRADE) and implement the corresponding race detection tool. Then, we introduce a new definition of transactional data race that is more intuitive, transparent to the underlying TM implementation, can be used for a broad set of C/C++ TM programs. Based on this new definition, we proposed T-Rex, an efficient and scalable race detection tool for C/C++ TM applications. Using TRADE and T-Rex, we have discovered subtle transactional data races in widely-used STAMP applications which have not been reported in the past.

The shift to multicore architectures will require new programming technologies that enable mainstream developers to write parallel programs that can safely take advantage of the parallelism offered by multicore processors. One challenging aspect of shared memory parallel programming is synchronization. Programmers have traditionally used locks for synchronization, but lock-based synchronization has well-known pitfalls that make it hard to use for building thread-safe and scalable software components. Memory transactions have emerged as a promising alternative to lock-based synchronization because they promise to eliminate many of the problems associated with locks. Transactional programming constructs, however, have overheads and require optimizations to make them practical. Transactions can also benefit significantly from hardware support, and multicore processors with their large transistor budgets and on-chip memory hierarchies have the opportunity to provide this support.

The advent of multi-core architectures has brought concurrent programming to the forefront of software development. In this context, Transactional Memory (TM) has gained increasing popularity as a simpler, attractive alternative to traditional lock-based synchronization. The recent integration of Hardware TM (HTM) in the last generation of Intel commodity processors turned TM into a mainstream technology, raising a number of questions on its future and that of concurrent programming. To evaluate the potential impact of Intel's HTM, we conducted the largest study on TM to date, comparing different locking techniques, hardware and software TMs, as well as different combinations of these mechanisms, from the dual perspective of performance and power consumption. As a result we perform a workload characterization, to help programmers better exploit the currently available TM facilities, and identify important research directions.

Chip multi-processors (CMPs) have become ubiquitous, while tools that ease concurrent programming have not. The promise of increased performance for all applications through ever more parallel hardware requires good tools for concurrent programming, especially for average programmers. Transactional memory (TM) has enjoyed recent interest as a tool that can help programmers program concurrently. The transactional memory (TM) research community is heavily invested in the claim that programming with transactional memory is easier than alternatives (like locks), but evidence for or against the veracity of this claim is scant. In this paper, we describe a user-study in which 237 undergraduate students in an operating systems course implement the same programs using coarse and fine-grain locks, monitors, and transactions. We surveyed the students after the assignment, and examined their code to determine the types and frequency of programming errors for each synchronization technique. Inexperienced programmers found baroque syntax a barrier to entry for transactional programming. On average, subjective evaluation showed that students found transactions harder to use than coarse-grain locks, but slightly easier to use than fine-grained locks. Detailed examination of synchronization errors in the students' code tells a rather different story. Overwhelmingly, the number and types of programming errors the students made was much lower for transactions than for locks. On a similar programming problem, over 70% of students made errors with fine-grained locking, while less than 10% made errors with transactions.

Transactional Memory (TM) is an emerging technology which simplifies the concurrency control in a parallel program. In this paper we propose Quick TM, a new hardware transactional memory (HTM) architecture. It incorporates three features to address known bottlenecks in the existing HTM architectures. First, we propose hardware-only dynamic detection of true-shared variables. Our result shows that true-shared variables account for only about 20% in the commit set of any transaction. Rest can be completely disregarded from the commit phase. This shortens every commit phase drastically resulting in a significant overall speed-up. Second, we keep both the speculative and the last committed version local to each processor. This benefits when a transaction is repeated in a loop. The processor request gets satisfied from the L1 data cache(L1D) itself. Furthermore, since both the versions are locally maintained, the commit action involves only broadcast of addresses. Third, we have proposed a mechanism to address overflow in transactions. In our proposal, each processor continues to run transactions even if one processor has overflown its L1D. Our technique eliminates the stall of a thread even if it conflicts with the overflown transaction. Overflown transaction commits in-place and periodically broadcasts its write set addresses, termed “partial commit”. This gradually reduces conflicts and allows other threads to progress towards commit. Moreover, the technique does not require any additional hardware at any memory hierarchy level beyond L1. Quick TM outperforms the state-of-the-art scalable HTM architecture, Scalable-TCC, on average by 20% in the latest TM benchmark suite STAMP. It outperforms the original TCC proposal with serialized commit by 28% on average. Maximum speed-up achieved in these two cases are 43% and 67% respectively. Our proposal handles transaction overflow gracefully and outperforms the current overflow-aware HTM proposal, One TM-concurrent by 12% on average.

Hardware transactional memory (HTM) offers a promising parallel programming model for chip multiprocessors. The performance aspect of HTMs has been explored extensively, whereas little research has addressed the power of HTMs. In this paper, we investigate power consumption in HTMs and propose two optimization techniques. The first optimization technique is adaptive snoop granularity (ASG). HTMs rely on cache coherence protocols to detect conflicts and maintain consistency of transactional data. One of the main design issues facing HTMs is the growing number of snoops required to maintain coherency of transactional data. We found that many transactions access consecutive memory locations which are not shared by other transactions. ASG monitors these transactional accesses and dynamically changes snoop granularity to reduce the power of the interconnection network and eliminate needless cache snoops. The second optimization technique is transactional snoop filtering (TSF). TSF dynamically tracks accesses to the coherence caches and eliminates cache snoops that would result in cache misses. TSF relies on small filters to monitor cache addresses. Energy is reduced as accesses to the much more demanding data caches are decreased. We extended the Gem5 simulator to model ASG and TSF. Our simulation results show that ASG and TSF are effective and reduce energy of interconnects and caches up to 44% and 89%, respectively.

Recent trends have led hardware manufacturers to place multiple processing cores on a single chip, making parallel programming the intended way of taking advantage of the increased processing power. However, bringing concurrency to average programmers is considered to be one of the major challenges in computer science today. The difficulty lies not only in writing correct parallel programs, but also in achieving the required efficiency and performance. For parallel programs, performance is not only about obtaining low execution times on a fixed number of cores, but also about maintaining efficiency as the number of available cores is increased. Ideally, programmers should have in their toolkit techniques that can be used when designing parallel programs, before any code is available for testing on production hardware, and which are then able to predict scalability once the program is implemented. Existing methods are either unreliable at predicting scalability, such as the case of disjoint-access parallelism, or do not apply to lock-based programs, which currently make up a large part of existing concurrent programs. Furthermore, using some of these techniques is so complicated that it outweighs the time required to implement, debug and test the program on real hardware. In this thesis we study the problem of predicting the scalability of concurrent programs without implementing them. This allows programmers in the design phase of a concurrent algorithm to choose only one or a few promising solutions that will be implemented, debugged and tested on production hardware. We first consider disjoint-access parallelism, an existing property that applies only to a very restricted class of programs. After an extensive practical evaluation spanning across a variety of scenarios, we find it to be ineffective at predicting scalability. For predicting the scalability of more general concurrent algorithms, we propose the obstruction degree, a new scalability metric based on the consistency requirements of algorithms. It applies to programs using locks, invalidation primitives and transactional memory. Our metric allows programmers to compare two given algorithms as well as predict their scalability limit, the maximum number of processors to which they can scale, thus allowing programmers to choose the appropriate size hardware for running their programs. We also examine the composition of relaxed memory transactions in order to combine the ease of programming offered by transactional memory with the increased scalability of transactions that circumvent the traditional transactional model. We present outheritance, a property we show to be both necessary and sufficient for ensuring the correct composition of relaxed transactions, and we show how to calculate the obstruction degree of compositions that use this new property. We use outheritance to build OE-STM, a new software transactional memory algorithm having elastic transactions that correctly compose.

This paper explores data speculation for improving the performance of Hardware Transactional Memory (HTM). We attempt to reduce transactional conflicts by decoupling them from cache coherence conflicts; many HTMs do not distinguish between transactional conflicts and coherence conflicts, leading to false transactional conflicts. We also attempt to mitigate the effects of coherence conflicts by using value prediction in transactions. We show that coherence decoupling and value prediction in transactions complement each other, because they both speculate on data in ways that are infeasible in the absence of HTM support.As a demonstration of how data speculation can improve performance, we introduce DPTM, a best-effort HTM that mitigates the effects of false sharing at the cache line level. DPTM does not alter the underlying cache coherence protocol, and requires only minor, processor-local, modifications.We evaluate DPTM against a baseline best-effort HTM, and compare it with data restructuring by padding, the most commonly used method to avoid false sharing. Our experiments show that DPTM can dramatically improve performance in the presence of false sharing without degrading performance in its absence, and consistently performs better than restructuring by padding.

Transactional memory (TM) has reached a maturity level and programmers have started using this programming model to parallelize their applications. However, although much effort has been put into the development of TM systems, there is still lack of debugging and development tools for TM applications, such as race detection tools.Previous definitions of transactional data race often impose constraints on the TM implementation or the programming language and cannot be widely applied to current STM designs. We propose a new definition of transactional data race that follows the programmer's intuition of racy accesses, is independent of thread interleaving, can accommodate popular STM systems, and allows common programming idioms.Based on this definition, we design and implement T-Rex, a precise dynamic race detection tool for C/C++ TM programs. Using T-Rex we discover transactional data races in STAMP applications that, to the best of our knowledge, have not been previously reported. Our experiments also show that T-Rex runtime overhead is comparable to state-of-the-art lock-based race detection tools, despite the extra work required to handle transactional memory semantics.

Executing Java programs with transactional memory

Computational Fluid Dynamics (CFD) is an increasingly important application domain for computational scientists. In this paper, we propose and analyze optimizations necessary to run CFD simulations consisting of multi-billion-cell mesh models on large processor systems. Our investigation leverages the general industrial Navier-Stokes open-source CFD application, Code_Saturne, developed by Electricité de France (EDF). Our work considers emerging processor features such as many-core, Symmetric Multi-threading (SMT), Single Instruction Multiple Data (SIMD), Transactional Memory, and Thread Level Speculation. Initially, we have targeted per-node performance improvements by reconstructing the code and data layouts to optimally use multiple threads. We present a general loop transformation that will enable the compiler to generate OpenMP threads effectively with minimal impact to overall code structure. A renumbering scheme for mesh faces is proposed to enhance thread-level parallelism and generally improve data locality. Performance results on IBM Blue Gene/P supercomputer and Intel Xeon Westmere cluster are included.

Hardware Transactional Memory (HTM) allows the use of transactions by programmers, making parallel programming easier and theoretically obtaining the performance of fine-grained locks. However, transactions can abort for a variety of reasons, resulting in the squash of speculatively executed instructions and the consequent loss in both performance and energy efficiency. Among the different sources of abort, conflicting concurrent accesses to the same shared memory locations from different transactions are often the prevalent cause.In this work, we characterize, for the first time to the best of our knowledge, how the aggressiveness of the cores in terms of exploiting instruction-level parallelism can interact with thread-level speculation support brought by HTM systems. We observe that altering the size of the structures that support out-of-order and speculative execution changes the number of aborts produced in the execution of transactional workloads on a best-effort HTM implementation. Our results show that a small number of powerful cores is more suitable for high-contention scenarios, whereas under low contention it is preferable to use a larger number of less aggressive cores. In addition, an aggressive core can lead to performance loss in medium-contention scenarios due to an increase in the number of aborts. We conclude that depending on contention, a careful choice over processor aggressiveness can reduce abort ratios.

Hardware Transactional Memory (HTM) gives software developers the opportunity to write parallel programs more easily compared to any previous programming method, and yields better performance than most previous lock-based synchronizations. Current implementations of HTM perform very well with small transactions. But when a transaction overflows the cache, these implementations either abort the transaction as unsuitable for HTM, and let software takeover, or revert to some much more inefficient hash-like in-memory structure, usually located in the userspace. We present a fast, scalable solution that has virtually no limit on transaction size, has low transactional read and write overhead, works with physical addresses, and doesn't require any changes inside the cache subsystem. This paper presents an HTMOS - Operating System (OS) and Architecture modifications that leverage the existing OS Virtual Memory mechanisms, to support unbounded transaction sizes, and provide transaction execution speed that does not decrease when transaction grows.

Performance evaluation of View-Oriented Transactional Memory

Hardware signatures have been used for detecting conflicts in Transactional Memory (TM) systems. Even with its area-efficiency, a signature can degrade TM performance by falsely detecting conflicts. Hence, increasing the accuracy of signatures with limited hardware resources is a crucial issue. We propose a simple and efficient signature design, unified signature. Instead of using separate read- and write-signatures, we implement a single signature to track all read- and write-accesses, effectively enlarging the signature size without additional overhead. Our results show that a TM system with a 4K-bit blind unified signature achieves average speedups of 8% over baseline TM systems with 38% less power and 17% less area.