Key History Tree: Efficient Group Key Management with Off-Line Members - Computer Network Security - page 250

5 Experimental Results

We have simulated extensively the use of a KHT within AGCD, described in the
previous section. We wanted to get a better understanding of how well the KHT
would behave when subject to a realistic workload. In particular, we wanted
to know what would be the size of the annotation that is required to obtain a
good approximation of the KHT-WS, and what is the benefit of KHT versus
just buffering aggregated LKH updates. Also, it was interesting to explore the
sensitivity of our results to different conditions, such as changing how often we
flush the group key or how clients behave while accessing the CDFE.

Register
Again LONG_TERM QUEUE

FAIL

1-P_doneLT

Access

Random

OK

Dlt

P_doneLT

Unregister

Register
Again

SHORT_TERM QUEUE

FAIL P_LT

P_doneST

Random

Register Access

Dst OK

1-(P_LT+P_doneST)

Fig. 4. Generating simulation traces

5.1 Simulation Methodology

Fig. 4 highlights a Jackson queuing network [13] that we use to generate simu-
lation traces of clients accessing an AGCD service. We distinguish between two
sets of clients: “short term” clients have “low commitment” with the service,
they are just starting to use it, and it is likely that a good proportion of them
will not be interested in the service after all. “Short term” clients that use the
service for a reasonable period of time become “long term”. We assume that
clients classified as “long term” are likely to keep using the service for a longer
time, but interaction with the service may be less frequent, i.e., there is no longer
a “novelty factor”.
The life-cycle of a client is to first register with the RFE and then, after a
delay modelled by an exponential distribution of mean D ST , start download-

ing the content from the CDFE. If decryption fails, he will have to re-register;
otherwise, we will assume that with certain probability

p doneST , he will aban-

don the service, or with probability p LT he will become a “long term” member;

otherwise, he will just retry the access after another random delay. Behaviour of
“long term” clients is similar but with different termination probability p doneLT ,