Analyse des performances de stockage, en mémoire et sur les périphériques d'entrée/sortie, à partir d'une trace d'exécution

Data storage is an essential resource for the computer industry. Storage devices must be fast and reliable to meet the growing demands of the data-driven economy. Storage technologies can be classified into two main categories: mass storage and main memory storage. Mass storage can store large amounts of data persistently. Data is saved locally on input/output devices, such as Hard Disk Drives (HDD) and Solid-State Drives (SSD), or remotely on distributed storage systems. Main memory storage temporarily holds the necessary data for running programs. Main memory is characterized by its high access speed, essential to quickly provide data to the Central Processing Unit (CPU). Operating systems use several mechanisms to manage storage devices, such as disk schedulers and memory allocators. The processing time of a storage request is affected by the interaction between several subsystems, which complicates the debugging task. Existing tools, such as benchmarking tools, provide a general idea of the overall system performance, but do not accurately identify the causes of poor performance. Dynamic analysis through execution tracing is a solution for the detailed runtime analysis of storage systems. Tracing collects precise data about the internal behavior of the system, which helps in detecting performance problems that are difficult to identify. The goal of this thesis is to provide a tool to analyze storage performance based on lowlevel trace events. The main challenges addressed by this tool are: collecting the required data using kernel and userspace tracing, limiting the overhead of tracing and the size of the generated traces, synchronizing the traces collected from different sources, providing multi-level analyses covering several aspects of storage performance, and lastly proposing abstractions allowing users to easily understand the traces. We carefully designed and inserted the instrumentation needed for the analyses. The tracepoints provide full visibility into the system and track the lifecycle of storage requests, from creation to processing. The Linux Trace Toolkit Next Generation (LTTng), a free and low-overhead tracer, is used for data collection. This tracer is characterized by its stability, and efficiency with highly parallel applications, thanks to the lock-free synchronization mechanisms used to update the content of the trace buffers. We also contributed to the creation of a patch that allows LTTng to capture the call stacks of userspace events.

Résumé

ABSTRACT: Data storage is an essential resource for the computer industry. Storage devices must be fast and reliable to meet the growing demands of the data-driven economy. Storage technologies can be classified into two main categories: mass storage and main memory storage. Mass storage can store large amounts of data persistently. Data is saved locally on input/output devices, such as Hard Disk Drives (HDD) and Solid-State Drives (SSD), or remotely on distributed storage systems. Main memory storage temporarily holds the necessary data for running programs. Main memory is characterized by its high access speed, essential to quickly provide data to the Central Processing Unit (CPU). Operating systems use several mechanisms to manage storage devices, such as disk schedulers and memory allocators. The processing time of a storage request is affected by the interaction between several subsystems, which complicates the debugging task. Existing tools, such as benchmarking tools, provide a general idea of the overall system performance, but do not accurately identify the causes of poor performance. Dynamic analysis through execution tracing is a solution for the detailed runtime analysis of storage systems. Tracing collects precise data about the internal behavior of the system, which helps in detecting performance problems that are difficult to identify. The goal of this thesis is to provide a tool to analyze storage performance based on lowlevel trace events. The main challenges addressed by this tool are: collecting the required data using kernel and userspace tracing, limiting the overhead of tracing and the size of the generated traces, synchronizing the traces collected from different sources, providing multi-level analyses covering several aspects of storage performance, and lastly proposing abstractions allowing users to easily understand the traces. We carefully designed and inserted the instrumentation needed for the analyses. The tracepoints provide full visibility into the system and track the lifecycle of storage requests, from creation to processing. The Linux Trace Toolkit Next Generation (LTTng), a free and low-overhead tracer, is used for data collection. This tracer is characterized by its stability, and efficiency with highly parallel applications, thanks to the lock-free synchronization mechanisms used to update the content of the trace buffers. We also contributed to the creation of a patch that allows LTTng to capture the call stacks of userspace events.