Endpoint Admission Control
WebTP Presentation
9/26/00
Presented by Ye Xia
Reference:
L. Breslau, E. W. Knightly, S. Shenkar, I. Stoica, H. Zhang,
“Endpoint Admission Control: Architectural Issues
And Performance”. Sigcomm 2001.
Why Endpoint?
• Aim of admission control (AC): provide QOS to
real-time flows
• IntServ has per-flow and router-based AC;
requires hop-by-hop signalling (RSVP); each
router keeps per-flow state; scalability problem.
• DiffServ lacks AC; providing QOS to each flow is
not a primary concern; but more scalable.
• Hope: endpoint AC can combine the strength of
both.
Algorithm
• Admission decision based on loss only
• Probing phase: each flow (at the end host)
probes the network for loss or marking ratio
(say, for 5 seconds)
• If the ratio is below a threshold, , flow is
admitted. i ri C
• Loss model:
r
i
i
Router scheduling mechanisms
• Fair Queueing has “stolen bandwidth” problem.
• Example: suppose two types of flows; r2 > r1; and
= 0.
• Type 1 flow is admitted if r1(n1+n2) < C; type 2
flow is admitted if r1n1 + r2n2 < C.
• When r1(n1+n2) = C, type 1 flows experience no
loss; type 2 flows’ loss ratio is (r2 – r1)/ r2
Best-Effort (TCP) Traffic
• Need to isolate TCP traffic and AC traffic.
Consider what happens when
– TCP traffic source is idle
– TCP induces loss
Architecture Choice
• Priority queues
– High priority for AC traffic
– Low priority for TCP traffic
– Probe traffic may take intermediate priority
– FIFO queueing for AC traffic
• AC traffic is rate-limited and served at that
rate.
– non-work conserving scheduler
Probing Algorithms
• Difficulty in sampling loss/mark ratio
• Out-of-band probing
– probing traffic takes lower priority than regular data traffic
– Probing traffic has higher loss
• ECN marking:
– marking rate higher than dropping rate
– Router simulates a virtual queue drained at 90% capacity
• Problem: cannot relate specified threshold, , with
actual loss ratio
Slow-Start Probing
• Thrashing: when many flows waiting for
admission, probing traffic overloads the
link.
• Cause: flow of rate r probes at rate r.
• Solution: slow-start probing. Gradually
ramp up rate of probing traffic.
Thrashing
• Utilization collapses for both in-band and out-band probing
• For in-band probing, data loss ratio increases as well
Simulation Models
• Leaky-bucket constrained traffic sources
– On-off sources and movie traces
• Poisson arrival of flows; exponential
holding time with mean 300s.
• Interfering TCP traffic needs not to be
simulated.
= 0, .01, .02, .03, .04, .05, .1, .15, .2.
• Comparison with router-based AC.
Traffic Sources
Basic Scenario
• Offered load: 20%
blocking prob.
• Loss rate competitive with
MBAC
is meaningful only for in-
band drop. Other probing
algo. reduce utilization.
• For in-band drop, 0.4%
loss rate when = 0.
• For out-band marking, low
loss ratio can be achieve
after probing for 5 seconds.
Longer Probing Time
• In-band dropping
• Lower loss ratio
and lower
utilization
High Load – In-band Dropping
• 400% offered
load; 75%
blocking prob.
• High loss
• Slow-start
probing does
better
High Load – Out-band Probing
• All algorithms are
similar
• Probing traffic does
not cause extra loss to
data traffic
• Slow-start probing has
higher utilization and
loss ratio
High Load - Marking
Heterogeneous Traffic
• Large flow has 4 times the peak rate
and higher blocking probability
• MBAC has similar behavoir
Multi-hop
Loss Probability
Multi-hop – Blocking Probability
Sharing FIFO Queue with TCP
• Two lower curves
are for = 0.04 and
0.05
• TCP prevents AC
traffic to be admitted
Comments
• Quick conclusion on queueing/scheduling
– Reconcile scheduling with end-to-end measurement
• Probing time is long.
– can aggregate probing traffic
– What to probe?
• AC criteria needs to be expanded (not just loss)
has no relationship with actual loss ratio
• WebTP has similar setup and similar issues.
