Search

We present Prompt Cache, an approach for accelerating inference for large language models (LLM) by reusing attention states across different LLM prompts. Many input prompts have overlapping text segments, such as system messages, prompt templates, and documents provided for context. Our key insight is that by precomputing and storing the attention states of these frequently occurring text segments on the inference server, we can efficiently reuse them when these segments appear in user prompts. Prompt Cache employs a schema to explicitly define such reusable text segments, called prompt modules. The schema ensures positional accuracy during attention state reuse and provides users with an interface to access cached states in their prompt. Using a prototype implementation, we evaluate Prompt Cache across several LLMs. We show that Prompt Cache significantly reduce latency in time-to-first-token, especially for longer prompts such as document-based question answering and recommendations. The improvements range from 8x for GPU-based inference to 60x for CPU-based inference, all while maintaining output accuracy and without the need for model parameter modifications., Comment: To appear at MLSys 2024

We consider the problem of fair resource allocation in a system where user demands are dynamic, that is, where user demands vary over time. Our key observation is that the classical max-min fairness algorithm for resource allocation provides many desirable properties (e.g., Pareto efficiency, strategy-proofness, and fairness), but only under the strong assumption of user demands being static over time. For the realistic case of dynamic user demands, the max-min fairness algorithm loses one or more of these properties. We present Karma, a new resource allocation mechanism for dynamic user demands. The key technical contribution in Karma is a credit-based resource allocation algorithm: in each quantum, users donate their unused resources and are assigned credits when other users borrow these resources; Karma carefully orchestrates the exchange of credits across users (based on their instantaneous demands, donated resources and borrowed resources), and performs prioritized resource allocation based on users' credits. We theoretically establish Karma guarantees related to Pareto efficiency, strategy-proofness, and fairness for dynamic user demands. Empirical evaluations over production workloads show that these properties translate well into practice: Karma is able to reduce disparity in performance across users to a bare minimum while maintaining Pareto-optimal system-wide performance., Comment: Full version of paper accepted to USENIX OSDI 2023 with proofs of theoretical guarantees

Caches at CPU nodes in disaggregated memory architectures amortize the high data access latency over the network. However, such caches are fundamentally unable to improve performance for workloads requiring pointer traversals across linked data structures. We argue for accelerating these pointer traversals closer to disaggregated memory in a manner that preserves expressiveness for supporting various linked structures, ensures energy efficiency and performance, and supports distributed execution. We design PULSE, a distributed pointer-traversal framework for rack-scale disaggregated memory to meet all the above requirements. Our evaluation of PULSE shows that it enables low-latency, high-throughput, and energy-efficient execution for a wide range of pointer traversal workloads on disaggregated memory that fare poorly with caching alone.

Hull is an accelerator-rich distributed implantable Brain-Computer Interface (BCI) that reads biological neurons at data rates that are 2-3 orders of magnitude higher than the prior state of art, while supporting many neuroscientific applications. Prior approaches have restricted brain interfacing to tens of megabits per second in order to meet two constraints necessary for effective operation and safe long-term implantation -- power dissipation under tens of milliwatts and response latencies in the tens of milliseconds. Hull also adheres to these constraints, but is able to interface with the brain at much higher data rates, thereby enabling, for the first time, BCI-driven research on and clinical treatment of brain-wide behaviors and diseases that require reading and stimulating many brain locations. Central to Hull's power efficiency is its realization as a distributed system of BCI nodes with accelerator-rich compute. Hull balances modular system layering with aggressive cross-layer hardware-software co-design to integrate compute, networking, and storage. The result is a lesson in designing networked distributed systems with hardware accelerators from the ground up., Comment: 16 pages, 13 figures

We explore the design of scalable synchronization primitives for disaggregated shared memory. Porting existing synchronization primitives to disaggregated shared memory results in poor scalability with the number of application threads because they layer synchronization primitives atop cache-coherence substrates, which engenders redundant inter-core communications. Substantially higher cache-coherence latency ($\mu$s) with substantially lower bandwidths in state-of-the-art disaggregated shared memory designs amplifies the impact of such redundant communications and precludes scalability. In this work, we argue for a co-design for the cache-coherence and synchronization layers for better performance scaling of multi-threaded applications on disaggregated memory. This is driven by our observation that synchronization primitives are essentially a generalization of cache-coherence protocols in time and space. We present GCS as an implementation of this co-design. GCS employs wait queues and arbitrarily-sized cache lines directly at the cache-coherence protocol layer for temporal and spatial generalization. We evaluate GCS against the layered approach for synchronization primitives: the pthread implementation of reader-writer lock, and show that GCS improves in-memory key-value store performance at scale by 1 - 2 orders of magnitude., Comment: 14 pages, 11 figures

Many applications that benefit from data offload to cloud services operate on private data. A now-long line of work has shown that, even when data is offloaded in an encrypted form, an adversary can learn sensitive information by analyzing data access patterns. Existing techniques for oblivious data access-that protect against access pattern attacks-require a centralized and stateful trusted proxy to orchestrate data accesses from applications to cloud services. We show that, in failure-prone deployments, such a centralized and stateful proxy results in violation of oblivious data access security guarantees and/or system unavailability. We thus initiate the study of distributed, fault-tolerant, oblivious data access. We present SHORTSTACK, a distributed proxy architecture for oblivious data access in failure-prone deployments. SHORTSTACK achieves the classical obliviousness guarantee--access patterns observed by the adversary being independent of the input--even under a powerful passive persistent adversary that can force failure of arbitrary (bounded-sized) subset of proxy servers at arbitrary times. We also introduce a security model that enables studying oblivious data access with distributed, failure-prone, servers. We provide a formal proof that SHORTSTACK enables oblivious data access under this model, and show empirically that SHORTSTACK performance scales near-linearly with number of distributed proxy servers., Comment: Full version of USENIX OSDI'22 paper

Memory-compute disaggregation promises transparent elasticity, high utilization and balanced usage for resources in data centers by physically separating memory and compute into network-attached resource "blades". However, existing designs achieve performance at the cost of resource elasticity, restricting memory sharing to a single compute blade to avoid costly memory coherence traffic over the network. In this work, we show that emerging programmable network switches can enable an efficient shared memory abstraction for disaggregated architectures by placing memory management logic in the network fabric. We find that centralizing memory management in the network permits bandwidth and latency-efficient realization of in-network cache coherence protocols, while programmable switch ASICs support other memory management logic at line-rate. We realize these insights into MIND, an in-network memory management unit for rack-scale memory disaggregation. MIND enables transparent resource elasticity while matching the performance of prior memory disaggregation proposals for real-world workloads., Comment: 18 pages, 9 figures, 2 tables

Serverless cloud computing handles virtually all the system administration operations needed to make it easier for programmers to use the cloud. It provides an interface that greatly simplifies cloud programming, and represents an evolution that parallels the transition from assembly language to high-level programming languages. This paper gives a quick history of cloud computing, including an accounting of the predictions of the 2009 Berkeley View of Cloud Computing paper, explains the motivation for serverless computing, describes applications that stretch the current limits of serverless, and then lists obstacles and research opportunities required for serverless computing to fulfill its full potential. Just as the 2009 paper identified challenges for the cloud and predicted they would be addressed and that cloud use would accelerate, we predict these issues are solvable and that serverless computing will grow to dominate the future of cloud computing.