Wesley Miaw

1
A Load Balancing Framework

• Wesley Miaw
• University of North Carolina at Chapel Hill

2
Load Balancing

• Move from big iron to many small cheap resources.
• Jobs distributed across a pool of resources.
• Performance minimize delay, maximize resource
  use/efficiency.

3
Load Balancing Applications

• Microprocessors Hyper-threading
• Web Servers EquiLoad
• Grid Computing Partitioning Dynamic Adaptive
  Grid Hierarchies (DAGP)

4
Grid Computing

• e.g. SETI_at_home
• Ad hoc, widely distributed, heterogeneous
  resources.
• Autonomous resource management.
• Client application submits independent
  computations for processing.

5
Framework

• Job Partitioning
• Job Distribution
• Resource Management
• Client Interface

6
Job Partitioning

• Classify jobs into groups by type and quantity of
  resources needed.
• Goal is to maximize performance when distributing
  jobs.

7
Hyper-threadPartitioning

• Each thread treated as a job.
• No job (thread) classification.
• Threads execute simultaneously on a single
  processor without context-switching.

8
EquiLoad Partitioning

• Intermediate server forwards requests over a
  cluster.
• Document size ltgt processing time.
• Create equal queues at each server.

9
EquiLoad Partitioning
10
DAGP Partitioning

• Grid computing maps resources onto a coordinate
  space.

11
DAGP Partitioning

• Partition the space to finish jobs as quickly as
  possible.

12
Framework

• Job Partitioning
• Job Distribution
• Resource Management
• Client Interface

13
Job Distribution

• Assign jobs to resources to maximize performance.
• Use partitioning information to make smart
  decisions.

14
Hyper-thread Distribution
15
EquiLoad Distribution

• HTTP Requests sent to dispatcher.
• Dispatcher forwards to correct host.
• Incorrect forwards redirected to correct host.

16
DAGP Distribution

• Expose data-parallelism.
• Minimize communication overhead.
• Balance load.
• Minimize repartitioning overhead.

17
Data-Parallelism

• Perform computations in parallel with respect to
  the data, not the computation.
• Perform the same operation on two halves of the
  data.
• Not performing different operations on the data
  simultaneously.

18
DAGP Distribution
19
Framework

• Job Partitioning
• Job Distribution
• Resource Management
• Client Interface

20
Resource Management

• Duplicated and additional resources need
  coordination.
• Manage resources to minimize communication,
  data-transfer, and shared resource overhead.

21
Hyper-thread Resources

• Shared resources memory bus, caches, execution
  units, branch predictors, queues.
• Duplicated resources TLBs, instruction pointers,
  processor state.
• Execution modes MT-mode, ST0-mode, ST1-mode.

22
EquiLoad Resources
23
DAGP Resources

• No special resource management.
• Grid computing is a resource management domain.

24
Framework

• Job Partitioning
• Job Distribution
• Resource Management
• Client Interface

25
Client Interface

• Simple interfaces are easier to use.
• Backwards-compatible interfaces can be deployed
  cheaply.

26
Hyper-thread Interface

• Hardware presents the illusion of two physical
  processors.
• Operating systems treat the hyper-threading
  processor as two physical processors.
• New HALT instruction transitions between MT-mode
  and STX-modes.

27
EquiLoad Interface

• Completely backwards-compatible with web
  browsers.
• Accept HTTP requests, return HTTP responses.
• A new interface would never be deployed.

28
DAGP Interface

• Identifying independent computation blocks is a
  manual process.
• Distribution of computation blocks may be handled
  by software.
• Use DAGP on the client to tell the grid where to
  distribute jobs.
• Use DAGP on the grid to distribute submitted jobs.

29
Framework

• Job Partitioning
• Job Distribution
• Resource Management
• Client Interface