Rack-Scale Memory Pooling for Datacenters

The rise of web-scale services has led to a staggering growth in user data on the Internet. To transform such a vast raw data into valuable information for the user and provide quality assurances, it is important to minimize access latency and enable in-memory processing. For more than a decade, the only practical way to accommodate for ever-growing data in memory has been to scale out server resources, which has led to the emergence of large-scale datacenters and distributed non-relational databases (NoSQL). Such horizontal scaling of resources translates to an increasing number of servers that participate in processing individual user requests. Typically, each user request results in hundreds of independent queries targeting different NoSQL nodes - servers, and the larger the number of servers involved, the higher the fan-out. To complete a single user request, all of the queries associated with that request have to complete first, and thus, the slowest query determines the completion time. Because of skewed popularity distributions and resource contention, the more servers we have, the harder it is to achieve high throughput and facilitate server utilization, without violating service level objectives. This thesis proposes rack-scale memory pooling (RSMP), a new scaling technique for future datacenters that reduces networking overheads and improves the performance of core datacenter software. RSMP is an approach to building larger, rack-scale capacity units for datacenters through specialized fabric interconnects with support for one-sided operations, and using them, in lieu of conventional servers (e.g. 1U), to scale out. We define an RSMP unit to be a server rack connecting 10s to 100s of servers to a secondary network enabling direct, low-latency access to the global memory of the rack. We, then, propose a new RSMP design - Scale-Out NUMA that leverages integration and a NUMA fabric to bridge the gap between local and remote memory to only 5× difference in access latency. Finally, we show how RSMP impacts NoSQL data serving, a key datacenter service used by most web-scale applications today. We show that using fewer larger data shards leads to less load imbalance and higher effective throughput, without violating applications¿ service level objectives. For example, by using Scale-Out NUMA, RSMP improves the throughput of a key-value store up to 8.2× over a traditional scale-out deployment.

The Case for RackOut: Scalable Data Serving Using Rack-Scale Systems

Edouard Bugnion, Alexandros Daglis, Babak Falsafi, Boris Robert Grot, Stanko Novakovic

To provide low latency and high throughput guarantees, most large key-value stores keep the data in the memory of many servers. Despite the natural parallelism across lookups, the load imbalance, introduced by heavy skew in the popularity distribution of keys, limits performance. To avoid violating tail latency service-level objectives, systems tend to keep server utilization low and organize the data in micro-shards, which provides units of migration and replication for the purpose of load balancing. These techniques reduce the skew, but incur additional monitoring, data replication and consistency maintenance overheads. In this work, we introduce RackOut, a memory pooling technique that leverages the one-sided remote read primitive of emerging rack-scale systems to mitigate load imbalance while respecting service-level objectives. In RackOut, the data is aggregated at rack-scale granularity, with all of the participating servers in the rack jointly servicing all of the rack’s micro-shards. We develop a queuing model to evaluate the impact of RackOut at the datacenter scale. In addition, we implement a RackOut proof-of-concept key-value store, evaluate it on two experimental platforms based on RDMA and Scale-Out NUMA, and use these results to validate the model. Our results show that RackOut can increase throughput up to 6× for RDMA and 8.6× for Scale-Out NUMA compared to a scale-out deployment, while respecting tight tail latency service-level objectives.

2016

Network-Compute Co-Design for Distributed In-Memory Computing

Alexandros Daglis

The booming popularity of online services is rapidly raising the demands for modern datacenters. In order to cope with data deluge, growing user bases, and tight quality of service constraints, service providers deploy massive datacenters with tens to hundreds of thousands of servers, keeping petabytes of latency-critical data memory resident. Such data distribution and the multi-tiered nature of the software used by feature-rich services results in frequent inter-server communication and remote memory access over the network. Hence, networking takes center stage in datacenters. In response to growing internal datacenter network traffic, networking technology is rapidly evolving. Lean user-level protocols, like RDMA, and high-performance fabrics have started making their appearance, dramatically reducing datacenter-wide network latency and offering unprecedented per-server bandwidth. At the same time, the end of Dennard scaling is grinding processor performance improvements to a halt. The net result is a growing mismatch between the per-server network and compute capabilities: it will soon be difficult for a server processor to utilize all of its available network bandwidth. Restoring balance between network and compute capabilities requires tighter co-design of the two. The network interface (NI) is of particular interest, as it lies on the boundary of network and compute. In this thesis, we focus on the design of an NI for a lightweight RDMA-like protocol and its full integration with modern manycore server processors. The NI capabilities scale with both the increasing network bandwidth and the growing number of cores on modern server processors. Leveraging our architecture's integrated NI logic, we introduce new functionality at the network endpoints that yields performance improvements for distributed systems. Such additions include new network operations with stronger semantics tailored to common application requirements and integrated logic for balancing network load across a modern processor's multiple cores. We make the case that exposing richer, end-to-end semantics to the NI is a unique enabler for optimizations that can reduce software complexity and remove significant load from the processor, contributing towards maintaining balance between the two valuable resources of network and compute. Overall, network-compute co-design is an approach that addresses challenges associated with the emerging technological mismatch of compute and networking capabilities, yielding significant performance improvements for distributed memory systems.

EPFL2018

Mitigating Load Imbalance in Distributed Data Serving with Rack-Scale Memory Pooling

Edouard Bugnion, Alexandros Daglis, Babak Falsafi, Boris Robert Grot, Stanko Novakovic, Dmitrii Ustiugov

To provide low-latency and high-throughput guarantees, most large key-value stores keep the data in the memory of many servers. Despite the natural parallelism across lookups, the load imbalance, introduced by heavy skew in the popularity distribution of keys, limits performance. To avoid violating tail latency service-level objectives, systems tend to keep server utilization low and organize the data in micro-shards, which provides units of migration and replication for the purpose of load balancing. These techniques reduce the skew but incur additional monitoring, data replication, and consistency maintenance overheads. In this work, we introduce RackOut, a memory pooling technique that leverages the one-sided remote read primitive of emerging rack-scale systems to mitigate load imbalance while respecting service-level objectives. In RackOut, the data are aggregated at rack-scale granularity, with all of the participating servers in the rack jointly servicing all of the rack’s micro-shards. We develop a queuing model to evaluate the impact of RackOut at the datacenter scale. In addition, we implement a RackOut proof-of-concept key value store, evaluate it on two experimental platforms based on RDMA and Scale-Out NUMA, and use these results to validate the model. We devise two distinct approaches to load balancing within a RackOut unit, one based on random selection of nodes—RackOut_static—and another one based on an adaptive load balancing mechanism— RackOut_adaptive. Our results show that RackOut_static increases throughput by up to 6× for RDMA and 8.6× for Scale-Out NUMA compared to a scale-out deployment, while respecting tight tail latency service-level objectives. RackOut_adaptive improves the throughput by 30% for workloads with 20% of writes over RackOut_static.