To match the blooming demand of generative AI workloads, GPU designers have so far been trying to pack more and more compute and memory into single complex and expensive packages. However, there is growing uncertainty about the scalability of individual GPUs and thus AI clusters, as state-of-the-art GPUs are already displaying packaging, yield, and cooling limitations. We propose to rethink the design and scaling of AI clusters through efficiently-connected large clusters of Lite-GPUs, GPUs with single, small dies and a fraction of the capabilities of larger GPUs. We think recent advances in co-packaged optics can enable distributing AI workloads onto many Lite-GPUs through high bandwidth and efficient communication. In this paper, we present the key benefits of Lite-GPUs on manufacturing cost, blast radius, yield, and power efficiency; and discuss systems opportunities and challenges around resource, workload, memory, and network management.

AI clusters today are one of the major uses of High Bandwidth Memory (HBM). However, HBM is suboptimal for AI workloads for several reasons. Analysis shows HBM is overprovisioned on write performance, but underprovisioned on density and read bandwidth, and also has significant energy per bit overheads. It is also expensive, with lower yield than DRAM due to manufacturing complexity. We propose a new memory class: Managed-Retention Memory (MRM), which is more optimized to store key data structures for AI inference workloads. We believe that MRM may finally provide a path to viability for technologies that were originally proposed to support Storage Class Memory (SCM). These technologies traditionally offered long-term persistence (10+ years) but provided poor IO performance and/or endurance. MRM makes different trade-offs, and by understanding the workload IO patterns, MRM foregoes long-term data retention and write performance for better potential performance on the metrics important for these workloads.

Sustainable and cost-effective long-term storage remains an unsolved problem. The most widely used storage technologies today are magnetic (hard disk drives and tape). They use media that degrades over time and has a limited lifetime, which leads to inefficient, wasteful, and costly solutions for long-lived data. This article presents Silica: the first cloud storage system for archival data underpinned by quartz glass, an extremely resilient media that allows data to be left in situ indefinitely. The hardware and software of Silica have been co-designed and co-optimized from the media up to the service level with sustainability as a primary objective. The design follows a cloud-first, data-driven methodology underpinned by principles derived from analyzing the archival workload of a large public cloud service. Silica can support a wide range of archival storage workloads and ushers in a new era of sustainable, cost-effective storage.

Automated storage and retrieval systems (ASRS) are a key component of the modern storage industry, and are used in a wide range of applications, carrying anything from lightweight tape cartridges to entire pallets of goods. Many of these systems are under pressure to maximise the use of space by growing in height and density, but this can create challenges for the the robots that service them. In this context, we present RASCAL, a novel ASRS robot for small payload items in structured environments, with a focus on system-level scalability and redundancy. We describe the design objectives of RASCAL and how they address some of the limitations of existing robotic systems in this area, such as scalability and redundancy. We then demonstrate the viability of our design with a proof-of-concept implementation of a data centre storage media robot, and show through a series of experiments that its design, speed, accuracy, and energy efficiency are appropriate for this application.

Sustainable and cost-effective long-term storage remains an unsolved problem. The most widely used storage technologies today are magnetic (hard disk drives and tape). They use media that degrades over time and has a limited lifetime, which leads to inefficient, wasteful, and costly solutions for long-lived data. This paper presents Silica: the first cloud storage system for archival data underpinned by quartz glass, an extremely resilient media that allows data to be left in situ indefinitely. The hardware and software of Silica have been co-designed and co-optimized from the media up to the service level with sustainability as a primary objective. The design follows a cloud-first, data-driven methodology underpinned by principles derived from analyzing the archival workload of a large public cloud service. Silica can support a wide range of archival storage workloads and ushers in a new era of sustainable, cost-effective storage.

DHT-based P2P systems provide a fault-tolerant and scalable means to store data blocks in a fully distributed way. Unfortunately, recent studies have shown that if connection/disconnection frequency is too high, data blocks may be lost. This is true for most of the current DHT-based systems' implementations. To deal with this problem, it is necessary to build more efficient replication and maintenance mechanisms. In this article, we study the effect of churn on PAST, an existing DHT-based P2P system. We then propose solutions to enhance churn tolerance and evaluate them through discrete event simulation.

Large-scale distributed systems gather thousands of peers spread all over the world. Such systems need to offer good routing performances regardless of their size and despite high churn rates. To achieve that requirement, the system must add appropriate shortcuts to its logical graph (overlay). However, to choose efficient shortcuts, peers need to obtain information about the overlay topology. In case of heterogeneous peer distributions, retrieving such information is not straightforward. Moreover, due to churn, the topology rapidly evolves, making gathered information obsolete. State of- the-art systems either avoid the problem by enforcing peers to adopt a uniform distribution or only partially fulfill these requirements. To cope with this problem, we propose DONUT, a mechanism to build a local map that approximates the peer distribution, allowing the peer to accurately estimate graph distance to other peers with a local algorithm. The evaluation performed with real latency and churn traces shows that our map increases the routing process efficiency by at least 20% compared to the state-of-the-art techniques. It points out that each map is lightweight and can be efficiently propagated through the network by consuming less than 10 bps on each peer.

Massively Multiplayer Online Games (MMOGs) recently emerged as a popular class of applications with millions of users. To offer acceptable gaming experience, such applications need to render the virtual world surrounding the player with a very low latency. However, current state-of-the-art MMOGs based on peer-to-peer overlays fail to satisfy these requirements. This happens because avatar mobility implies many data exchanges through the overlay. As state-of-the-art overlays do not anticipate this mobility, the needed data is not delivered on time, which leads to transient failures at the application level. To solve this problem, we propose Blue Banana, a mechanism that models and predicts avatar movement, allowing the overlay to adapt itself by anticipation to the MMOG needs. Our evaluation is based on large-scale traces derived from Second life. It shows that our anticipation mechanism decreases by 20% the number of transient failures with only a network overhead of 2%.