Cache Copy-On-Write: Research Of Mechanism Of Speculated Data Management In TLS

With chip integration continues to improve, more and more processor cores are integrated into a single chip, which results in the emergence of many-core processors. Different with the rapid development of multi-core architecture, parallel programming did not achieve a leapfrog development. Although traditional parallel programming models, such as data-parallel programming model, message passing programming model, shared memory programming model, make the programmers’ work of parallelization easier to some extent, but they lack of a certain support for some coarse-grained parallelism and they do not support the thread speculative execution, which also causes a huge waste of computing resources.TLS(Thread-Level Speculation) is one of the typical mechanism in terms of mining coarse-grained parallelism in applications in the way of speculation.Its biggest advantage is the very simple programming model. The programmer can simply identify those sections of code that can be speculatively executed and the runtime system or hardware are responsible for ensuring data dependencies between speculative threads are not destroyed. However, because of its high overhead and other reasons, TLS mechanism is not widely used.This paper presents and implements a kind of speculative data management mechanisms for TLS system, which has a simple programming model and low-overhead system management. This mechanism not only implements an efficient speculative data management, under the circumstance of no reduction in the clock frequency of processor core, and to some extent, solves the problem of sharing data hazard between speculative threads.The works and results of this paper are:1 The proposed concept of life cycle of the speculative variableThe life cycle of speculative variable includes the generation, use and submission of speculative variable. According to the life cycle of speculative variable, this paper proposes a new TLS technical classification method and compares the advantages and disadvantages of existing methods. In the meantime, this paper sums up the multi-core platform design space, which supports TLS, and proposes several ways to explore the design space. At last, this paper notes the current challenges and future trends in TLS.2 The propose of Cache Copy-on-Write speculative data management mechanism based on the life cycle of speculative variableBecause there is no special optimization for speculative data management in traditional TLS system, this results in its high-overhead and poor performance. Based on the TLS technical analysis, this paper proposes Cache Copy-on-Write speculative data management based on the life cycle of speculative data. Cache Copy-on-Write mechanism implements dynamic management to speculative data, which is similar to Copy on Write technology in operating system, and eliminates the overhead of copying data when threads are built. In addition, Cache Copy-on-Write mechanism supports speculative thread communication, through which TLS can eliminate many thread rollbacks due to sharing data hazard between speculative threads. Because of this, the successful rate of thread execution is enhanced and so does the performance.3 The implement of speculative data management mechanism using Enhanced Local MemoryFrom the point of performance, Cache Copy-on-Write mechanism has the ability of high-efficiency speculative data management and low management overhead. However, using the Enhanced Local Memory as speculative thread private data cache can maximize Cache Copy-on-Write mechanism advantages. First, Enhanced Local Memory has the characteristics of high speed. Secondly, Enhanced Local Memory is the on-chip memory. Based on these two points and combined with Cache Copy-on-Write mechanism, this paper implements a highly efficient data management TLS system.Based on SESC simulator, this paper implements a TLS speculative data management mechanism using Enhanced Local Memory and carries out a detailed analysis of performance testing. The experimental results demonstrate that the proposed mechanism for speculative data management is efficient and effective.