This is our automated approach to generating optimized execution plans for serverless DAGs, with regards to latency objective or budget.
Appeared in Sigmetrics Conference, 2022
WISEFUSE has the following three key optimizations:
It combines in-series functions together within one VM to reduce overhead communication between cascaded functions.
It executes a group of parallel invocations of a function within a VM, improving resource sharing.
It improves resource allocation, assigning the correct VM size to each function/function bundle to reduce E2E latency and cost.
Orion
This is our offering for estimating E2E latency and allocating resources to meet probabilistic guarantees for E2E latency. Compatible with AWS Lambda.
Appeared in OSDI, 2022
Orion has the following three distinctive features:
It creates a performance model accounting for runtime dependencies within a DAG.
It uses a method for co-locating parallel invocations in a VM, mitigating skew among the invocations.
It uses a method for pre-warming VMs for ensuing functions in a DAG.
SONIC
This is our offering for serverless applications and customizes how to pass data between serverless functions to minimize latency. Compatible with AWS Lambda.
SONIC is a data-passing manager that optimizes application performance and cost, by transparently selecting the optimal data-passing method for each edge of a serverless workflow DAG and implementing communication-aware function
placement.
Appeared in Usenix Annual Technical Conference, 2021
Sonic has the following three distinctive features:
It applies to the emerging class of serverless applications.
It dynamically selects the best data passing options balancing latency and cost.
It does not need any modification to the applications.
OptimusCloud
This is our dynamic, cloud-hosted database tuner, optimized for Cassandra and AWS.
Appeared in Usenix Annual Technical Conference, 2020. US Patent Pending.
OptimusCloud has the following three distinctive features:
It considers heterogeneous cluster configurations.
It jointly optimizes database and VM configurations.
It dynamically adjusts configuration as workload behavior changes.
Sophia
This is our dynamic, on-premise database tuner for more privacy-minded users, optimized for Cassandra and Redis and benchmarked against ScyllaDB.
Appeared in Usenix Annual Technical Conference, 2019. US Patent Pending.
Sophia has the following three distinctive features:
It performs cost-benefit analysis to achieve long-horizon optimized performance for clustered NoSQL instances in the face of dynamic workload changes.
It executes a decentralized protocol to gracefully switch over the cluster to the new configuration while respecting data consistency requirements.
It maintains data availability throughout the reconfiguration.
Rafiki
This is our static database tuner. Given a static workload and execution environment, it provides the optimal configuration for your Cassandra database.
Appeared in ACM/IFIP/USENIX Middleware Conference, 2017
