Information Technology Reference
In-Depth Information
and average delay between accesses
D
LT
. Client arrivals are characterized by a
Poisson process with
λ
arrivals per day on average.
# Evi
c
tions
# Active Clients
200000
100000
Wo
rkloads
Wo
rkloads
New
New
150000
80000
Stable
Stable
60000
100000
40000
50000
20000
100
Days
100
Days
20
40
60
80
20
40
60
80
Fig. 5.
Simulator input traces workloads
Fig. 5 shows characteristics of two different sets of workloads that we will
use in our evaluation. One of them tries to model a “new” service that has a
very significant number of “short term” clients, i.e., half of them on average. The
other one is a “stable” or “well-established” service where a large proportion of
clients are “long term”. In both cases we adjust
so that the growth in average
number of active members mimics the declared progress of a real and widely used
content delivery subscription service [14]. Clearly, the eviction rate of the “new”
service is much higher than the “well-established” one. Table 1 gives details of
the configuration parameters used.
λ
Table 1.
Workload configuration parameters
Workload
p
doneST
p
LT
p
doneLT
D
ST
D
LT
λ
Stable
0.02
0.18
0.02
1
2
1800
New
0.1
0.1
0.02
1
2
3000
In order to compute confidence intervals we used five different traces of each
set that used non-overlapping random number sequences with the same seed.
This was possible because the high quality random number generator (RNG)
that we used had a period long enough to accommodate all these sequences.
Results presented use a 0.95 confidence intervals margin.
5.2
KHT versus Buffered LKH
We show in Fig. 6 how well KHT compares to a buffered version of LKH. This
use of LKH accumulates all the group additions and evictions since the last key
refresh into a highly optimized key update. Somehow, this is equivalent to using
algorithms that optimize multiple additions or evictions in LKH [1], rather than
the more inecient single change counterparts.
