PASSTA (IPC and Synchronizing Shared data on heterogeneous MPSoCs) is a project in cooperation with Huawei focusing on assisting traditional Operating System services with hardware. Currently, we investigate Inter-Process Communication (IPC) mechanisms in Linux and plan to extend the developed concepts further to applocations within micro-kernels.
IPC is a general term for mechanisms used to communicate amongst different processes. This communication can be used for data transfer, synchronization, or both. Traditionally, IPC functionalities are tightly integrated into the Operating System as they are highly critical for achieving good throughputs and latencies. The ongoing development towards heterogeneous multi-/many-core processor architectures exposes performance insufficiencies in established IPC services. An increasing number of CPUs and more fine-grained multi-threaded/-process applications lead to more dependencies and data exchange between different application parts, utilizing IPC services to synchronize accesses to the same dataset. The ever-increasing utilization of IPC mechanisms highlights these insufficiencies, which impact the applications' performance.
When multiple threads interact, scenarios can occur where a thread must wait for a certain condition before continuing its execution. For instance, such a condition may be a free lock or the availability of data. An event can be considered as a state change of this condition and can be triggered by different sources such as device drivers, communication channels or interacting threads. Event notification is used to inform that a particular event of interest has occurred. In PASSTA, we focus on blocking event notification mechanisms. In these mechanisms, the event-receiving thread calls a specific function to get new events. This function returns successfully if an event is available. Otherwise, this function puts the thread asleep, waiting for the corresponding condition to be met (blocking behavior). An event-generating thread has to notify the sleeping thread about a condition change and thus an occurred event by triggering its wake-up (event notification).
To improve this event notification we develop in PASSTA a concept to assist blocking IPC mechansisms with a hardware component (HWAcc). When improving event notification, two metrics have to considered:
- Metric 1 - Syscall duration: As depicted in the figure with M1, the syscall includes the event generation and the event notification that initiates the thread wake-up. The CPU cycles required for the syscall show the overhead that an event notification can add to a thread that generates an event.
- Metric 2 - Wake-up latency: This metric denotes the number of cycles for a thread wake-up initiation, and is labeled with M2 in the figure. For this, we measure the time from the start of the event notification function in thread B until thread A is active. As the HWAcc processes a request asynchronously, only metric 2 includes the execution time of the HWAcc itself.
Reduce burden for event-generating thread (Metric 1):
Thread A has to wait for a particular condition before it can continue its execution, e.g., if a lock is not available (Cond. Check). In Linux, to be notified about a change in this condition, a wait list is used to specify the expected notification when this event occurs. This wait list is filled by thread A before it goes to sleep. Notification of a sleeping thread is initiated by its wake-up. Many wait lists exist in the kernel, each tied to a certain element (e.g., a file descriptor), while the wait list structure is always the same. After thread B generates an event the condition becomes valid, e.g., if a lock is released (Event Gen.). Therefore thread B checks whether another thread was waiting for the event by querying the wait list, and to wake up thread A in the original event notification approach (Event Notify).
In PASSTA we developed a concept to facilitate the event notification by initiating the thread wake-up from a hardware unit (HWAcc), thus relieving the thread that generates an event. This can be achieved by replacing the default notification function in the wait list in step Cond. Check with one that offloads the wake-up initiation to the HWAcc. The HWAcc then asynchronously initiates a thread wake-up.
Reducing the latency in blocking IPC mechanisms (Metric 2):
Several steps are involved in the wake-up procedure triggered by an event-generating thread. First, the waker (thread B) determines which thread to wake up and where to wake it up. After that, an IRQ/IPI is sent to the core on which thread A should be woken up. On the wakee side in the interrupt service routine, the wakelist is checked, and consequently, the scheduler is triggered to determine whether the newly awakened thread A should run on the core. All these steps contribute to the latency of IPC and consist mainly of scheduling-related functions.
We aim to decrease the time spent in scheduling-related functions to reduce the latency of blocking IPC mechanisms. Therefore, we introduce a hardware-assisted scheduling class into Linux, which offloads scheduling functionalities to a dedicated hardware unit. This newly created scheduling class coexists with the standard scheduling classes, thus enabling a seamless integration into the Linux kernel.