Buffer Management for Shared-Memory ATM Switches

1
Buffer Management for Shared-Memory ATM Switches

• Written By Mutlu Apraci John A.Copelan
• Georgia Institute of Technology
• Presented By Yan Huang

2
Outline

• Describe several buffer management policies, and
  their strengths and weakness.
• Evaluate the performance of various policies
  using computer simulations
• Compare of the most import schemes

3
Some basic definition

• The prime purpose of an ATM switch is to route
  incoming cells arriving on a particular input
  link to the output link (switching)
• Three basic techniques are used
• space-division crossbar switch
• shared-mediumbased on a common high-speed bus
• shared-memory
• Also have the functionality of queue
• Input queuing, output queuing, shared memory

4
Shared-Memory Switch

• Consists of a single dual-ported memory shared by
  all input and output line
• Bring both switching and queuing
• Does not suffer from the throughput
• degradation caused by head of the line
  blocking(HOL)
• The main focus is buffer allocation
• determines how the total buffer spaces(memory)
  will be used by individual output ports of the
  switches. (Contd)

5
Shared-Memory Switch (Contd)

• The selection and implementation of the buffer
  allocation policy is refereed as buffer
  management
• Model of the SM
• switch
• N output ports
• M buffer space
• Performance
• cell loss
• occurs when a cell
• arrive at a switch
• node and find the buffer is full

6
Buffer Allocation Policies

• Stochastic assumption
• Poisson arrivals
• Exponential service time
• Static Thresholds
• Complete Partition
• The entire buffer space is permanently
  partitioned among the N servers.
• Does not provide any sharing
• Complete Sharing
• An arriving packet is accepted if any space is
  available in the switch memory
• Independent of the server to which the packet is
  directed

7
Comparison of CP and CS

• CP policy
• the buffer allocated to a port is wasted if that
  port is inactive, since it can not be used by
  other possibly active lines
• CS policy
• one of the ports may monopolize most of the
  storage space if it is highly utilized.
• In the CS policy, a packet is lost
• when the common memory is full.
• In CP, a packet is lost when its
• corresponding queue has already
• reaches its maximum allocation.
• The assumption of the traffic
• arrival process enable us to model
• the switch as a Markov
• process (Fig 3)

8
Simulation

• The assumption of exponential inter arrival and
  service time dist is not realistic for ATM system
• The traffic in ATM networks is bursty in nature.
• To model it, use an ON/OFF source
• Simulation
• mean duration of ON state 240.
• Mean duration of OFF state 720
• cell interarrival time 5
• Switch model has two output ports(N2)
• The size of the shared memory is 300 cells(M300)
• Performance metric is the cell loss ration(CLR)
  at the port

9
Performance of CS and CP

• Balanced traffic load at the port are equal
• For medium traffic load, CS achieve lower CLR

10
Performance of CS and CP(contd)

• Imbalanced traffic
• The load at one port is varied, but remain
  constant at the other port
• CS both port have the same CLR
• CP port buffer are isolated. CLR at port 1
  increase with the traffic load

11
Sharing with Maximum Queue Length

• SMXQ -a limit is imposed on the number of buffers
  to be allocated at any time to any server.
• There is one global threshold for all the queues.
  The advantage of
• SMXQ
• SMXQ achieves lower CLR than CP,
• manages to isolate the goodport
• from the bad port.
• The better CLR performance is
• obtained with buffer sharing,
• the isolation is obtained by restricting
• the queue length.

12
SMA and SMQMA

• Two variation of SMXQ
• SMA (sharing with a minimum allocation)
• A minimum number of buffer is always reserved for
  each port.
• SMQMA (sharing with a maximum queue and minimum
  allocation )
• each port always has access to a minimum
  allocated space, but they cannot have arbitrarily
  long queues.
• SMQMA has the following advantage over SMXQ
• A minimum space is allocated for each port in
  order to simplify the issue of serving
  high-priority traffic in a buffer-sharing
  environment.

13
Push-Out

• Push-out (PO) drop-on-demand(DoD)
• A previously accepted packet can be dropped from
  the longest queue in the switch
• Advantage
• Fair.
• Efficient.
• Naturally adaptive.
• Achieves a lower CLR than
• the optimal SMXQ setting
• Drawback
• Difficult to implement

14
Push-Out with Threshold

• In ATM networks, different ports carrying
  different traffic type might have different
  priorities.
• A modification to PO, CSVP is to achieve
  priorities among ports. Similar idea is called
  POT (push-out with threshold)
• CSVP has the following attributes.
• N users share the total available buffer space M,
  which is virtually partitioned into N segments
  corresponding to the N ports
• When the buffer is full, there are two
  possibilities
• If the arriving cells type, i , occupies less
  space than its allocation Ki, then, at least one
  other type must be occupy more than its own
  allocation, for instance, Kj. The admission
  policy will admit the newly arriving type i cell
  by pushing out a type j cell.
• If the arriving cells queue exceeds its
  allocation at the time of arrival, then the cell
  will be rejected.
• When the buffer is not full
• CSVP operates as CS.
• Under heavy traffic loads, the system tends to
  become a CP management.

15
Dynamic Policies

• The analyses of the buffer allocation problem
  above assume static environments
• Dynamic Threshold (DT)can be used to adapt to
  changes in traffic conditions.
• The queue length thresholds of the ports, are
  proportional to the current amount of unused
  buffering in the switch. T(t) a (M- Q(t))
• Cell arrivals for an output port are blocked
  whenever the output ports queue length equals or
  exceed the current threshold value
• Major advantage of DT is to be its robustness to
  traffic load changes, a feature not present in
  the ST policy

16
Comparison
17
The End