Resource Management for Scalability and Performance

1
Resource Management for Scalability and
Performance

• By Christopher Stanley
• Date 9-26-00

2
Chapter Goal

• Learn techniques to improve the size and
  performance of Net-VEs by reducing
• Network bandwidth
• Processor resources

3
Networked Virtual Environment Information
Principle

• Resources M x H x B x T x P
• M of messages transmitted in the net-VE
• H avg of destination hosts for each message
• B avg amount of bandwidth required for msg
• T packet delivery requirements (Large T
  less delay, small T larger delay)
• P of processor cycles required to receive and
  process each message

4
Resources to Manage

• Communication protocol optimization
• Data flow restriction
• Leveraging of limited user perception
• System architecture modifications

5
Communication Protocol Steps

• Allocated buffer for packet
• Write data to buffer
• Include packet header of protocol (ie- TCP/IP,
  UDP/IP, or IP multicast)
• Transmitted a packet containing the buffer
  contents and header

6
Optimizing the Communications Protocol

• Packet Compression reduces the size of each
  packet
• Packet Aggregation reduces the number of
  packets sent by packet merging
• Controlling data visibility
• Multicasting

7
Packet Compression Types

• Internal
• a. Lossless changes encoding format and
  eliminate redundant data in packet
• b. Lossy eliminates data that is irrelevant or
  reduce the detail of packet
• External
• a. Lossless Avoid retransmitting packets
  identical to previous packets
• b. Lossy Avoid retransmitting packets similar
  to previous packets
• 3. Protocol Independent Compression Algorithm
  (PICA)

8
PICA

• Eliminates redundant state info from successive
  update packets by
• 1. Transmitting a numbered state snapshot
• 2. Packets indicates which state snap shot it is
  referring to, and what states in that snap shot
  have changed

9
Application Gateway

• 1. Selects when a protocol codec should be used.
• 2. Detects when an entity has become inactive or
  dead
• 3. Filters updates of inactive entities from
  being sent over the WAN
• 4. Provides QES for inactive entities.

10
Packet Aggregation

• Saves bandwidth by reducing the number of packet
  headers transmitted.
• Header sizes UDP/IP 28 bytes, TCP/IP 40 bytes
  
• Could artificially delay packet transmission

11
Packet Aggregation Methods

• Timeout Based guarantees an upper bound to the
  delay place on packets
• Quorum Based guarantees a particular bandwidth
  and packet rate reduction
• Note Trades off packet delivery delay with
  bandwidth savings

12
Aggregation Servers

• Benefits
• Distributes workload across multiple processors
• Improved fault tolerance
• Can improve overall performance

13
Controlling the Visibility of Data

• Goal
• Decrease of H and B by increasing M.
• Assumption
• A user only needs a small portion of the total
  Net-VE data

14
Data flow management methods

• Area-of-interest provide Net-VE a data filter
  so that the VE only sends the host what it needs
• Multicasting restricts data flow by
  subscription-base network routing.
• Subscription-based aggregation Hybrid of 1 2

15
When to use filtering subscriptions

• Hosts have limited processor capacity that cant
  be wasted processing unwanted packets.
• Hosts have extremely low-bandwidth so network
  optimization is a must.
• Multicast/broadcast protocols arent available.
• Client subscription patterns change too rapidly
  for multicast regrouping to stay below max
  desired delay.
• Net-VE data isnt categorize able or a group per
  entity is too resource expensive.

16
Multicasting

• Setting up
• Multicast group setup at an multicast address
• Host subscribe/unsubscribe to the multicast group
• Any host can transmit data to multicast group
  regardless of subscription status
• Challenges
• Partitioning data across the multicast groups so
  subscribers will be interested in all or most of
  data
• Worst case
• each client is interested in data from all
  multicast groups multicast degenerates into a
  broadcast system

17
Subscription Based Aggregation

• Group-per-Entity Allocation
• Each entity is assigned a multicast group
• Each host can then subscribe to the groups or
  entities of interest
• Must provide an entity directory service
• Group-per-Region Allocation
• Entities subscribe to regions of interest

18
Group-per-Entity Allocation

• Problems
• Network traffic increases from retrieving the
  entities address list, and maintaining groups.
• Network cards limit the number of simultaneous
  multicast address subscriptions to 12, before
  running in software packet processing mode
• Using multiple multicast addresses increases the
  chances of network collisions
• Each multicast group should only have one sender
  to keep network router stack sizes small and to
  optimizes the multicast packet processing.

19
Group-per-Region Allocation

• Requires each entity to change its target
  group(s) as it travels throughout the virtual
  world
• Must track the entitys bounds of its current
  region
• Find the multicast addresses associated with the
  next region it enters
• A hexagonal grid is used to lessen the number of
  intersecting groups at a cross-over

20
Hybrid Multicast Aggregation

• Goal
• Find balance between fine-grain data
  partitioning and multicast grouping.

21
Three Tiered Interest Management

• Filters
• Group-per-Region data is segmented by location
• Group-per-Entity allows receivers to select
  individual entities at a fine-grain level of
  detail
• Area-of-Interest Filter subscriptions are
  supported by library routines
• Projections
• Are multi-attribute multicast groups
• Projections restrict the values of variables in
  the entitys packets

22
Communication Protocol Headers

• http//www.protocols.com/pbook/tcpip.htm