Análise do impacto de diferentes versionamentos de dados das memórias transacionais sobre memórias Phase-Change

Abstract

Two of the major issues in current large computer systems are energy consumption and how to explore concurrent systems in a correct and efficient way. Phase-Change Memories and Transactional Memories are two technologies that intend to solve these issues. Phase-Change Memory (PCM) is a new memory technology being studied to re-place DRAMs as the main memory in large data centers, as its non-volatility reduces static power consumption. The main problem of PCMs consists in its write operations, which are slow and generate degradation in material, thus reducing its life. Transactional memories are synchronization methods developed to reduce the dif-ficulties and limitations of lock-based methods. Their main advantages are related to being high-level and allowing composition and reuse of code. Another advantage of transactional memories compared to locks is the absence of deadlocks. Transactional memories are based on database transactions. Transactions in database systems meet four properties: atomicity, consistency, isolation and durability, or ACID. Transac-tional memories must also implement the ACID properties, except for durability. Trans-actional memories implement version management of data to ensure atomicity. The objective of this study is to analyze the impact on the PCM of different version management techniques implemented by STMs. To that end, the Phase-ChangeMem-ory - Multicore Simulator (PCM-MS) was implemented, a memory hierarchy simulator for multi-core systems where the PCM is the main memory. It determines changed bits in PCM to estimate the wear and energy consumption. In addition to the PCM-MS, Pintools was used to generate trace files that run in the simulator. As the STM library, TinySTM was chosen because it implement various version management and it rep-resents the state-of-art in STM. As benchmarks, Eigenbench and the STAMP set of benchmarks were used. The results showed that the WBC VM had the lowest wear on the PCM in 3 of 7 benchmarks analyzed. These results are related to the number of aborts of VMs, where the WBC presents a much smaller number of aborts than others VM, being up to 39 times lower in the experiment with the benchmark Kmeans with 64 threads. In future works, we intend to enhance the simulator and make the impact analysis in PCM of others transactional systems.