Multi-tenant GPU clusters are designed to concurrently run multiple distributed ML training workloads. However, frequent data transfers among GPUs via collective operations can slow down training, as collectives from different tenants compete for network bandwidth. Recent work (e.g., CASSINI) has considered collective coordination to prevent network contention, but primarily focused on static job-level optimizations at deployment time, oblivious to runtime network conditions and the specific traffic pattern of each workload. In this paper, we present Symphony, an application-layer solution that dynamically coordinates collective operations across tenants at runtime. Symphony integrates seamlessly with existing clusters with minimal modifications to the collective communication library and includes a lightweight online scheduling mechanism that requires no advance information about the workloads or their collectives. We evaluate Symphony using both a real GPU testbed implementation and trace-driven simulations. Specifically, using realistic ML workloads in our testbed, we observe improvements of up to 13.2% in average communication time and 9.6% in training time compared to state-of-the-art solutions.

Next generation applications for datacenters, such as Distributed Machine Learning (DML) and Big Data, have complex communication patterns that demand a scalable, stateless and application-aware optimal transport protocol to maximize network utilization and improve application performance. Recent transport protocols either provide limited benefits due to lack of information sharing between application and network; or implement complex stateful mechanisms to improve the application performance. In this paper, we present Omni- Transport Mechanism (Omni-TM) as a message-based congestion control protocol. Omni-TM allows exchanging message information with the network to negotiate the optimal transmission rate without maintaining a per-flow state at the switches (i.e., stateless). Omni- Tmis designed to reach maximum link capacity in one shot. Our simulation results show that Omni- Tmdemonstrates better traffic control decisions (i.e., close to zero queue length while maintaining high link utilization). Furthermore, Omni- Tmreduces Flow Completion Time (FCT) up to 45 % in a realistic workload compared to DCTCP.

It is common for storage systems designed to run on edge datacenters to avoid the high latencies associated with geo-distribution by relying on eventually consistent models to replicate data. Eventual consistency works well for many edge applications because as long as the client interacts with the same replica, the storage system can provide session consistency, a stronger consistency model that has two additional important properties: (i) read-your-writes, where subsequent reads by a client that has updated an object will return the updated value or a newer one; and, (ii) monotonic reads, where if a client has seen a particular value for an object, subsequent reads will return the same value or a newer one. While session consistency does not guarantee that different clients will perceive updates in the same order, it nevertheless presents each individual client with an intuitive view of the world that is consistent with the client's own actions. Unfortunately, these consistency guarantees break down when a client interacts with multiple replicas housed on different datacenters over time, either as a result of application partitioning, or client or code mobility. SessionStore is a datastore for fog/edge computing that ensures session consistency on a top of otherwise eventually consistent replicas. SessionStore enforces session consistency by grouping related data accesses into a session, and using a session-aware reconciliation algorithm to reconcile only the data that is relevant to the session when switching between replicas. This approach reduces data transfer and latency by up to 90% compared to full replica reconciliation.

In many edge computing scenarios data is generated over a wide geographic area and is stored near the edges, before being pushed upstream to a hierarchy of data centers. Querying such geo-distributed data traditionally falls into two general approaches: push incoming queries down to the edge where the data is, or run them locally in the cloud. Feather is a hybrid querying scheme that exploits the hierarchical structure of such geo-distributed systems to trade temporal accuracy (freshness) for improved latency and reduced bandwidth. Rather than pushing queries to the edge or executing them in the cloud, Feather selectively pushes queries towards the edge while guaranteeing a user-supplied per-query freshness limit. Partial results are then aggregated along the path to the cloud, until a final result is provided with guaranteed freshness. We evaluate Feather in controlled experiments using real-world geo-tagged traces, as well as a real system running across 10 datacenters in 3 continents. Feather combines the best of cloud and edge execution, answering queries with a fraction of edge latency, providing fresher answers than cloud, while reducing network bandwidth and load on edges.

No abstract available.