Runtimes. Serverless computing has appealing characteristics, based on simplicity, high scalability and fine grained execution. It has seduced both industry [10, 107, 118] and academia [64]. This enthusiasm has also led to a blossom of open-source systems (e.g., [1–4, 64] to cite a few). At core, a cloud function runtime is in charge of maintaining the user-defined functions, execut- ing them upon request. It must ensure strong isolation between function instances and low startup latency for performance. Many works propose to tackle these two central challenges. Micro-kernels [100] offer a solid basis to quickly kick-off a function and attain sub-millisecond startup time. Catalyser introduces the sfork system call to reuse the state of a running sandbox. Similarly, Firecracker [7] makes containers more lightweight and faster to spawn. SOCK [111] is a serverless-specialized system that uses a provisioning mechanism to cache and clone function con- tainers. SAND [8] exploits function interaction in XxxX to improve the performance applications. The system relaxes isolation at the application level, enabling functions from the same application to share memory and communicate through a hierarchical message bus. This allows better latency and resource efficiency when combining functions. Faasm [137] offers similar guarantees using a language-agnostic runtime built atop WebAssembly. User functions use all the same substrate to ex- ecute, allowing fast initialization. They can access a distributed key-value store cached locally and shared across functions located on the same physical machine. Two recent works [63, 75] coincide with our view that existing runtimes do not support mutable shared state and coordination across cloud functions. Xxxxxxxxxxx et al. [63] underline that the model is System Shared state Synchronization Durability Consistency PyWren Object store coarse-grained replication weak ExCamera rendezvous rendezvous — — Ripple composition composition — — Pocket/Crail multi-tiered coarse-grained ephemeral — Cloudburst XxxX + cache coarse-grained replication weak Faasm shared memory fine-grained — —- Crucial DSO fine-grained replication strong Table 8: Serverless solutions for state sharing and coordination. a data-shipping architecture that imposes indirect communication and hinders coordination. Xxxxx et al. [75] highlight the lack of adequate storage for fine-grained operations and the inability to coor- dinate functions at fine granularity.
Runtimes. All of these items are collectively referred to herein as “Materials”. If not marked or marked “None” in Schedule A, then such Materials are not licensed by this Agreement.
Runtimes. 6.1 Services that are consumed by the Customer shall be measured as a unit of Active and Static Runtimes.
