MemX virtualizes cluster-wide memory into a shared, low-latency resource pool for virtualized environments, enabling applications to access datasets larger than the memory capacity of a single machine. By aggregating distributed memory resources and providing transparent remote memory access, MemX allows cloud providers and data center operators to run large memory-intensive workloads with significantly lower latency and improved memory utilization.
Data-intensive applications running in virtualized environments are limited by the amount of physical memory and the latency of disk-based storage. Many workloads that involve large datasets, AI model training and inference, web services, and data analytics require access to datasets that exceed the memory capacity of a single machine. Existing solutions do not work well with standard virtualized environments. A new solution is needed to provide scalable, low-latency memory access in distributed systems while remaining transparent to existing applications running in virtualized environments.
MemX is a distributed system that aggregates cluster-wide memory and presents it as virtualized memory to VMs, enabling low-latency access to remote memory resources. The system uses a client-server architecture with a custom remote memory access protocol to enable low-latency, high-throughput communication between nodes. MemX also incorporates page-level memory de-duplication to reduce overall memory consumption across the cluster. The system is fully transparent to existing applications, requires no changes to software, libraries, or operating systems, while supporting live VM migration across nodes.
![]()
• Reduces random read latency from ~9 ms (virtual disk) to ~160 µs (MemX)
• Reduces sequential read latency from ~9 ms to under 120 µs
• Improves overall cluster memory efficiency through page-level de-duplication of identical memory content
• Achieves up to 85% memory savings via local page de-duplication in VM creation workloads
• Requires no application modifications for deployment
• Requires no recompilation or specialized APIs for integration
• Fully transparent to existing applications, libraries, and operating systems
• Supports live VM migration across cluster nodes
• Cloud computing platforms for running large-memory virtual machine workloads, such as large databases, data mining systems, and real-time analytics pipelines, that require low-latency data access
• High-performance computing clusters that require scalable memory beyond single-node physical limitations
• Enterprise virtualization environments for improving server consolidation and resource utilization efficiency
• AI model training and inference environments handling large datasets
• Distributed data analytics and web service infrastructures requiring high-performance memory access
• United States Patent 8,935,506 – Utility Patent, Filed 3/22/2012, Issued 1/13/2015, Patented
• IEEE publication available:
https://ieeexplore.ieee.org/document/5599239
Prototype – Implemented and benchmarked on a Xen/KVM/Linux cluster testbed. TRL ~5.
This technology is available for licensing.
Strong potential for cloud service providers, virtualization platform vendors, enterprise data centers, and high-performance computing organizations seeking scalable memory architectures that improve application performance while reducing infrastructure costs and increasing resource utilization.
Prototype benchmarking results, virtualization deployment details, and performance evaluation data are available upon request.