Memory Management for Scalable Web Data Servers

1
Memory Management for Scalable Web Data Servers

• S. Venkataraman, M. Livny, J. F. Naughton

2
Motivation

• Popular Web sites have heavy traffic
• 2 main bottlenecks
• CPU overhead of servicing TCP/IP connections
• I/O of fetching the many pages serviced
• CPU solution cluster of servers connected to one
  File Server
• If each node had own subset of web site, the
  cluster is succeptible to skew.

3
Motivation (continued)

• Papers Goal Develop buffer management
  techniques to best utilize the aggregate memory
  of the machines in the cluster to reduce I/O
  bottleneck.

4
Outline

• Web Server Architecture
• 3 Memory Management Techniques
• Client-Server
• Duplicate Elimination
• Hybrid
• Performance Evaluation
• Discussion

5
Web Server Architecture

• Cluster of cooperating servers connected by a
  fast network
• Each node has own disks and memory
• Each node runs the same copy of the server
• Round robin router distributes requests

6
Web Server Architecture Part Deux

• Primary server Server where the client request
  is serviced.
• Owner Server that manages a persistent data
  item.
• Owners maintain a directory of copies of the data
  pages they own in global memory.
• Paper considers algorithms for read-only
  workloads.

7
Memory Management

• Memory Hierarchy
• Primary server memory
• Owner memory
• memory at other servers
• disk
• Each request is broken into page-sized units.
• If Primary has page in memory, then done.
  Otherwise, ask Owner for it.

8
More on Memory Management

• Owner gets another node to forward page if
  possible.
• Otherwise, it gets it from disk and keeps a copy
  in memory.
• The page in Owner memory is labeled as hated.
• Eviction Policy hated pages are evicted first.
• When primary server receives a page, it must
  choose another page to evict. Three algorithms
  (no hated pages)
• Client-Server
• Duplicate Elimination
• Hybrid

9
Client-Server

• An LRU list of pages is kept.
• Very simple.
• Increases local hits.
• Lots of duplication possible.

10
Duplicate Elimination

• Considers the cost difference between evicting a
  single page (singlet) and a duplicate page.
• Duplicate pages eliminated first since they are
  cheep to fetch.
• 2 LRU lists singlets and duplicates
• Increases the percentage of database in memory.
• Main Drawback hot duplicate page is replaced
  before a cold singlet.
• But how do we keep track of duplicates?

11
Keeping Track of Duplicates

• When a page goes from singlet to duplicate
• Happens during a forward, so this is trivial. (no
  additional messages)
• When a page goes from duplicate to singlet
• Owner receives a message that a page was evicted.
• If only one copy remains, owner sends a message
  to that server.
• message can be piggy backed. (no additional
  messages)

12
Hybrids are Our Friends

• Estimate performance impact of eviction on the
  next page reference.
• Consider both likelihood of reference and cost of
  re-access.
• Latency of fetching page p back into memory
  C(p).
• cost of going to disk vs. cost of going to remote
  memory.
• Likelihood that page p will be accessed next
  W(p)
• W(p) 1 / (Elapsed time since last reference)
• Expected cost E(p) W(p) C(p)

13
More on Hybrid Algorithm

• Two LRU lists maintained just like Duplicate
  Elimination
• At eviction, heads of the two lists are compared.
• Replace page with lower Expected Cost

14
Simulation Model and Workload

• 8 nodes of 64 MB each.
• Message cost 1ms/page
• Link Bandwidth 15MB / sec
• All files are the same size.
• Access frequency Zipfian distribution.
• zipf parameter controls wide range of skews
• zipf of zero means access frequencies are uniform
• Pages of files are declustered among all servers.

15
And the survey says . . .

• Duplication is good at high skews, but bad at low
  skews
• At low skews (uniform access frequency)
• Duplicate elimination has good global memroy
  utilization.
• At high skews (zipf over 1)
• Client-Server keeps hot pages at all nodes in
  memory
• duplication is good
• Hybrid nears the performance of the better choice
  in both scenarios.
• Varying database size, using diverse file sizes,
  and more nodes gave similar results.

16
Discussion

• Web sites have predictable hit rates, can that be
  used somehow?
• Can we recycle evicted pages?