MemExchange: Cloud-Scale Memory Trading

MemExchange: Utility-Driven Distributed Memory Reallocation for Multi-Tenant Datacenters

To handle unpredictable workloads, cloud providers typically over-provision memory to meet peak demand, resulting in substantial underutilization across datacenter clusters. At the same time, memory-constrained tenants may suffer elevated cache miss rates, even when idle capacity remains stranded elsewhere in the infrastructure. MemExchange is a cluster-wide, multi-tenant memory management system that dynamically right-sizes in-memory caching tenants according to workload demand. Leveraging marginal-utility-based allocation derived from online Miss Ratio Curve (MRC) estimation, MemExchange redistributes idle memory between tenants across physical nodes using RDMA. This approach transforms the dedicated caching memory scattered across servers into a logically aggregated pool, enabling cross-node memory exchange without centralized coordination or forced tenant co-location. To support efficient remote access, we design the MemExchange Tracker Communication (MTC) protocol, an application-layer mechanism that coordinates memory reallocation and enables one-sided RDMA operations without involving remote CPUs. We implement MemExchange in Memcached and evaluate it through microbenchmarks, medium and rack-scale deployments of up to 100 CloudLab servers. Our results show up to 2.3x lower remote-access overhead compared to TCP-based designs, a 13% increase in cluster-wide memory utilization at rack scale, and up to 63% reduction in miss rate for memory-constrained tenants under skewed workloads.

FATE: Fairness-Aware Memory Allocation for Multi-Tenant Tiered Memory Environments

Hardware-Transparent I/O Governance in Disaggregated Heterogeneous Storage

Shared-nothing disaggregated storage clusters that serve both latency-sensitive databases and opaque block-volume workloads face two governance problems unsolved by existing schedulers: maintaining consistent performance across heterogeneous hardware generations, and enforcing global I/O limits when access patterns skew to a subset of storage nodes. We present the I/O Resource Manager (IORM), a multi-stage distributed scheduler deployed in production within Oracle Exadata Exascale. IORM combines three mechanisms: a hardware-aware cost modeler that normalizes I/O accounting using datasheet-derived fixed costs to make limits invariant across hardware generations; a quantum-based rate limiter with bounded carry-forward credits that accommodates database micro-bursts while enforcing long-term SLOs; and a distributed adaptive feedback controller that redistributes unused entitlements across the cluster to resolve topological access skew. Beyond design, we share operational lessons from production deployment. On an 8-node test cluster running up to 100 concurrent tenant volumes, IORM converges within 5\% of provisioned limits under extreme sequential skew, scales without inter-tenant interference, and recovers full throughput within 15 seconds of a storage-node failure.

Performance Prediction of Data-Rebalancing Actions in Sharded NoSQL Clusters

Scalable storage systems typically provide mechanisms for re-distributing data over time to ensure a balanced storage allocation. Such mechanisms have traditionally been considered as background activities meant to run at low priority, to avoid penalizing applications accessing storage. In this work, we focus on an alternative design point where data re-distribution is run as a high-priority activity, meant to re-balance data at the full speed allowed by the newly added resources. Such rapid redistribution has become viable in recent years due to the rise of low-overhead technologies in the networking and disk storage space. A challenge in such a scenario is to schedule data transfer flows to newly added nodes in a way that fully and efficiently utilizes network, CPU, and disk resources in the new nodes, while avoiding overload. To understand the impact of different parameters of the rebalance process, such as the number of simultaneous senders, we develop a queueing network (QN) model of the process used in MongoDB, and describe a preliminary evaluation of performance prediction by the QN model via simulations.

Scaling Unmodified Multithreaded Applications with Elastic CXL-based Distributed Shared Memory

While CXL presents a promising hardware substrate for Distributed Shared Memory (DSM), seamlessly scaling multithreaded applications across multiple nodes remains a formidable challenge. Existing CXL-based DSMs fall short: they require manual code modifications to share non-heap data, employ rigid data placement policies that fail under diverse and dynamic workloads, and suffer from severe page-fault processing overheads in sub-microsecond ($\mu\mathrm{s}$) environments. We present xDSM, a full-space, elastic DSM system built over CXL that transparently scales unmodified multithreaded applications. To eliminate the burden of manual code rewrites, xDSM employs an OS-runtime co-design that establishes a globally coordinated address space, seamlessly sharing all memory segments. To mask CXL access penalties, xDSM abandons static placement rules in favor of a dynamic, latency-driven policy that actively balances data between local DRAM and CXL memory. Finally, to resolve the fundamental tension between high base-page fault overheads and severe huge-page false sharing, xDSM introduces spatial locality-aware elasticity, dynamically coalescing and splitting pages on the fly to amortize processing costs. Evaluated across diverse workloads using 15 system configurations, xDSM outperforms CXL-only baselines by 1.5$\times$ to 2.2$\times$ and state-of-the-art hybrid DSMs by 1.1$\times$ to 2.2$\times$, while achieving near-linear scalability.

Preserving Admission Responsibility in Multi-Tenant Large Language Model Prefix Caches

Shared prefix caching turns Graphics Processing Unit (GPU) memory into persistent state shared across Large Language Model (LLM) tenants. A group that materializes new Key-Value (KV) blocks can force another to lose reusable state, yet request-time schedulers account for transient service, replacement policies primarily rank object value, and static partitioning strands idle capacity. We call this mismatch the admission-responsibility gap. To close it, we propose PrefixShield, which meters newly materialized full KV blocks, carries responsibility across requests, gates reuse promotion while debt remains, and uses projected debt to select the group supplying eviction candidates. We implement PrefixShield in vLLM. In paired runs under one-touch pollution, PrefixShield improves victim cache hit ratio by 9.39 percentage points over the Least Recently Used (LRU) policy and 8.64 points over S3-FIFO, restoring the victim from 4.92% to 84.87% at 4096-block scale, and gains 2.00 points over S3-FIFO under two-pass replay. It preserves benign ShareGPT behavior and work-conserving access to idle capacity. Delayed replay yields a 35.16-point advantage while debt remains. These results show that object-value signals rank what to retain, while persistent responsibility determines which group bears reclamation pressure.