CAPITULO I MARCO REFERENCIAL
MARCO TEÓRICO
3.2. Empleo, pleno empleo, inflación y bienestar.
First, the performance of PERM’s hybrid scheduler was compared against several existing multihoming schedulers from the literature. To this end, both hash-based and load-balancing scheduling algorithms [63] were implemented as well as sliding window active probing [24]. Hashing and load balancing are simple, light-weight approaches which do not use any information about the underlying links. To improve schedul-
Scheduler Mean Transmission Time in sec. Light Volume Heavy Volume
PERM (Min-Total) 1.50487 186.21
PERM (Min-Max) - 199.14
Active Probing 1.71785 201.23
Load Balancing 2.04534 211.37
Hashing 2.07801 257.57
Table 6.2: Mean transmission times for light- and heavy- volume flows under different schedulers during controlled experi- ments. 0% 20% 40% 60% 80% 100% Hash Load Probing Min-Max Min-Tot
Average Utilization of Total Capacity
Scheduler
Avg. DSL usage Avg. Cable usage
Figure 6.6: Average utilization of total available capacity from both access links combined during heavy-volume controlled experiments.
ing performance, the links can be probed to profile link characteristics allowing the scheduler to make more accurate decisions. The results confirm that there is a performance improvement of probing-based approaches over simpler approaches. The experiments further demonstrate that PERM’s predictive mech- anisms lead to higher performance than other probing-based schedulers.
Light-volume Flows
To quantify the benefits coming from PERM’s more accurate probing, the transfer times of small web con- nections were measured with different multihoming schedulers. The experiment was run using traces de- rived from several of the busiest users in the Dartmouth traces not used in the traffic analysis from Sec- tion 6.3.2. Because IP addresses in the original traces are sanitized, the IP addresses in the traces were replaced with the IP addresses of the top 500 web-sites as listed in the end-user based web-site rankings from the Alexa.com web information archive. The original access pattern from the traces was unchanged. For each scheduler the index page from each web site in the trace was downloaded. Multiple iterations were conducted for each scheduler at different times of the day. The first column of Table 6.2 shows the overall mean transmission time for each scheduler. The PERM scheduler has the lowest overall mean trans-
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 -40% -30% -20% -10% 0% 10% 20% 30% 40% 50% Improvement in Transfer Time
over Cable over DSL
(a) CDF of Improvement in Light-Volume Flow- Transmission Times 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 20 40 60 80 100 120 140 160 180 200 Per-flow Throughput in Kbps Cable DSL PERM
(b) CDF of Heavy-Volume Per-Flow Throughputs
Figure 6.7: Controlled Experiments: (a) Percent improvement in transmission times for light-volume flows. (b) Per-flow throughput for bulk transfers.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 500 1000 1500 2000 2500 3000 Per-flow Throughput in Kbps Cable DSL PERM
Figure 6.8: CDF of Per-flow throughput for upload flows.
mission time, 38%, 35%, and 14% lower than the other schedulers. Considering that light volume flows take up more that 60% of all connections, these results confirm that there is noticeable benefit from active probing in general and even more so from PERM’s predictive probing.
Heavy-volume Flows
The transmission time of heavy-volume flows is constrained more by available bandwidth rather than RTT. Because small flows do not compete for bandwidth with the large flows, to evaluate the effectiveness of the PERM scheduler for managing the available bandwidth several large files from the Internet were downloaded simultaneously. In the Dartmouth traces, 50% of users averaged only 1.25 or fewer simulta- neous large-volume flows of more than 20KB. 96% of users averaged 4 or fewer simultaneous flows. Of the large-volume flows, 92% were less than 1MB although transfers of over 2GB were also present. Files of slightly larger size are used for the experiment to ensure that downloads would sufficiently compete with
one another and stress the system. Files were hosted on different web-servers ranging in size from 1.7MB to 30MB covering 99% of all transfers in the traces.
To test the schedulers, several different iterations of downloading 4 different files simultaneously were performed. The downloads combined saturated the underlying links. The two PERM heavy-volume sched- ulers described in Section 6.3.3 are compared against the other multihoming algorithms. The mean total transfer times over 15 experimental iterations from each scheduler are shown in the second column of Ta- ble 6.2. The two PERM schedulers outperform all other schedulers although the Min-Max scheduler was only slightly better than the Adaptive Probing scheduler. The Min-Total scheduler improves the over- all download times by 27.7%, 11.9% and 7.4% over each of the non-PERM schedulers. More insight into the relative performance of each scheduler can be gained by examining the utilization of each access link. Figure 6.6 plots the average utilization of the total combined bandwidth available during the duration of the experiment. The usage of each individual link is shown as two separate parts of the total utilization. Again, the two PERM schedulers were able to make most use of the total capacity. Moreover, the Cable with its larger bandwidth is favored by both PERM schedulers. Although the DSL was under-utilized (38% for Min-Max and 65% for Min-Total scheduler), the utilization of the combined total capacity is highest.