GMS

Introduction

• Cluster-wide management
• CPU, high speed N/W (compared to disk)
• OS integrated, not exposed to App
• Workload: Idle resources

Methods

When page out, push to global memory before page out to disk. Every node has both local and global pages.

Global LRU

Decided by global age.

There is too much overhead to maintain a global age. Instead, a probablistic algorithm is used to decide which one to page out.

$w = \frac { T} { M}$

M pages to evict, T is on each node, how many fall into the oldest M.

$w_1, w_2, \cdots, w_j$ are used as weights to choose page out place. The node with more old pages have greater chance to be the destination.

Scale

Around 100s. Not work well when too many nodes in a cluster. To solve, break a big cluster down to smaller ones.

Take-aways

Ideal: Optimal solution Real: Roll the dice, get the sub-optimal solution, but keep the entire expectation good.