Projects

Properties of FIFO and CLOCK cache under general renewal traffic

We extend Che’s approximation to analyze CLOCK cache performance under both IRM and Hyperexponential Renewal request models and address three key properties: (1) IRM always yields well-behaved, convex LRU miss ratio curves; (2) with a fixed inter-arrival time distribution, FIFO performs worse when item popularity is skewed; conversely, (3) CLOCK performs better when item popularity is skewed. We present formal proofs for these properties.

Configurable and Cache-Accurate Trace Generation for Storage Benchmarking

We introduce 2DIO, a two-dimensional trace generation framework that encodes cache behaviors through a compact, quantized representation of recency patterns (IRD), combined with an independent reference model that characterizes frequency. This approach requires minimal parameters yet accurately reproduces complex, non-convex LRU miss ratio curves observed in real workloads.

Dynamic, Hierarchical Resource Allocation with Throttling for vSAN

This project focuses on performance isolation in VMware vSAN, using a dynamic linear programming model to optimize traffic allocations across virtual and physical disks while subject to both SLA and device capacity constraints. This approach takes advantage of the EXSi hypervisor's capabilities to throttle individual VMs, avoid heavy global data movement, and effectively isolate burst traffic.

A Distributed Object Store built on Conventional File System

This is an emulation of a distributed store system built on Zoned Namespace SSDs. The system guarantees strong consistency and fault tolerant.

Emulating an ISC-enabled LSM-tree for Read Optimization

This project presents a proof of concept for an ISC-enabled LSM-tree, aiming to enhance read performance and space efficiency by offloading parallel search computations from the host to storage devices.

A simple cache simulator lib

We introduce an LRU 2C-array algorithm; it is an efficient method for implementing a cache eviction policy using a fixed-size array; it is particularly suitable for scenarios where memory constraints prevent the use of linked lists.

A reproduction of the GAN I/O generation model

This is a reproduction of the GAN IO generation model proposed by the paper 'Accurate Generation of I/O Workloads Using Generative Adversarial Networks'.