Chapter 3 Behaviors and performance analysis of load-balancing algorithms
3.4 Performance analysis
3.4.1 Aggregation analysis of load-balancing parameters
Let LM be a link with the bandwidth which can take the sum of many narrower link ’s bandwidth. The objective function of bandwidth aggregation can be represented as (11). The goal is to use the most economical and reliable combination of existing
links to realize the bandwidth demand of LM. Different countries might have various combinations to meet this cost and bandwidth requirements since could be 256kbps in England while 128kbps in the U.S. However, our focus is not to find the combinations, but instead we want to make sure that it is possible to gain the throughput (LM) from a combination of many with a lower cost.
l
l
l
For practical reasons, we pick one of the available combinations from the market. By surveying many ISPs we find that using one T1 leased line and three lines of 512k/512 k ADSL links is a proper choice. (See table 2)
23
Table 2 T1 and 512k ADSL monthly fees for the three major ISPs in Taiwan:
ISP Name T1 monthly fee
(NT$)
512k ADSL monthly fee
(NT$)
CHT 54,600 3,700
SPARQ 145,000 3,700
TFN 144,000 3,700
Based on the discussion in section 3.1, in a connection-based system the LLB algorithm employs a Measuring-Type to process traffic load measurement at a distance D for every Mτ. The measurement results are evaluated in order to assign a BL by the Dispatching-scheme.
We can compare the performance and utilization of the generic parameters of the LLB algorithms. Moreover, the impact of link numbers on the bandwidth aggregation will be considered. Note that 6 lines of 256K are also included to see if the number of lines is sensitive to the experiment. The workload starts from Ln(3,1) to Ln(10,3).
Ln(3,1) is about the capacity of the bandwidth provided by a T1 leased line. To focus don the bandwidth aggregation issue, this simulation environment only has traffic generated from the testing clients, and no disturbance of outward traffic is induced.
3.4.1.1 Comparison on path-selecting period
Figure 5 (a) shows that when the workload reaches L3 (3, 3), the mean throughput of 3 aggregated links is getting closer to one T1 link. Note that in Figure 5 (a), the mean throughput of an IP-based dispatching period (PSip ) in every workload is
24
0 200 400 600 800
L1(3,1) L1(3,3) L1(10,3) L3(3,1) L3(3,3) L3(10,3) L6(3,1) L6(3,3) L6(10,3) Link & Traffic type
L1(3,1) L1(3,3) L1(10,3) L3(3,1) L3(3,3) L3(10,3) L6(3,1) L6(3,3) L6(10,3) Link & Traffic type
Figure 5 (a) Comparing mean throughput of path selection periods: connection-based (PSc) and IP-based (PSip) (b) EBU of the two path selection periods
worse than the connection-based dispatching period (PSc). Moreover, due to flow control mechanism to be explained later in this section, the throughput of 3 balanced links (L3’s) is always better than L6’s.
In Figure 5(b) the EBU values of PSc are all higher than those of PSip. Note that EBUs in Figure 5(b) in fact correlate to the mean throughput in Figure 5(a). An interesting situation is that the EBU value of L1 (3,1) is higher than L1(3,3). To compute the EBU value of L1(3,1) and L1(3,3) by using Equation (10), the total
25
service rate , the total bandwidth required ,
and the testing complete time (T
∑
= it is the time needed to complete the data transfer and depends on flow control behavior. When the arrival rate of traffic is higher than the service rate of the link, there will be packet losses and retransmissions. TCP flow control will start to lower down the end-to-end transmitting speed and make Tc longer.Like the mean throughput in Figure 5(a), Figure 5(b) also shows that the EBU of 3 balanced links is better than their counterpart of 6 links. The average measurement of Tc of L6(UN,SN) is longer than L3(UN,SN), which indicates that the flow control will have more impact on the narrower bandwidth.
3.4.1.2 Comparison of dispatching scheme and measuring period Figure 6(a) compares the mean throughput of the weighed dispatching scheme versus the best dispatching scheme with different measuring times, T = 1 and T= 0.1 seconds. When the workload is small, the best dispatching scheme (T= 0.1 sec) gets the best result. When the workloads are heavier, Ln(3,3) and Ln(10,3), the weighted
26
dispatching scheme outperforms the ‘best’ and its throughput approaches the T1 rate .
L1(3,1) L1(3,3) L1(10,3) L3(3,1) L3(3,3) L3(10,3) L6(3,1) L6(3,3) L6(10,3)
Link & Traffic type
L1(3,1) L1(3,3) L1(10,3) L3(3,1) L3(3,3) L3(10,3) L6(3,1) L6(3,3) L6(10,3) Link & Traffic type
Figure 6 (a) Comparing mean throughput of dispatching scheme and measuring time T (b) EBU of dispatching schemes and measuring times
The weighted approach achieves good throughput performance only when the number of sessions is high enough so that their load can be distributed. In the comparison of measuring time T, (T= 1 sec) is always worse than (T=0.1 sec), as this is a direct consequence of a smaller ∆Tjps-m (see Equation (4)) since the measurement
27
error can be reduced accordingly. Figure 6(b) shows that utilization correlates to mean throughput in Figure 6(a).
3.4.1.3 Comparison of measuring type and measuring distance
L1(3,1) L1(3,3) L1(10,3) L3(3,1) L3(3,3) L3(10,3) L6(3,1) L6(3,3) L6(10,3) Link & Traffic type
L1(3,1) L1(3,3) L1(10,3) L3(3,1) L3(3,3) L3(10,3) L6(3,1) L6(3,3) L6(10,3) Link & Traffic type
Figure 7 (a) Comparing mean throughput of measuring types and distance (b) EBU of the three measuring types
28
Figure 7(a) shows the results of various measuring types and the corresponding measuring distance as discussed in section 3.1.5. When the workload is small, the mean throughput is higher in the CSN and MRT types of algorithms. When the workload is heavier, the ST-type algorithm gets the higher throughput. However, there are no discriminations between good and bad L6 situations.
Figure 7(b) shows that the CSN type gets the highest EBU. To discuss this phenomenon, we should understand more about the operations of CSN-type algorithms. CSN-type algorithms use the counted number of sessions of each BLi to determine the path for transmissions. The path assignment operation of the CSN-type algorithm can be expressed by Equation (6) by setting each Wi =1. This usage of the weight utility function indicates that the CSN-type algorithms have a built-in weight property. This built-in weight property enforces that the CSN-type algorithms have the weighted dispatching behavior which can gain a higher utilization of bandwidth.
Thus, the result of Figure 7(b) of a higher EBU value with the CSN-type algorithm is due to the weighted dispatching property.
3.4.1.4 Summary of experiment results
To summarize up, the important factors that influence the bandwidth aggregation performance are the path-selecting period, the measuring period, and the dispatching scheme. Either the measuring type or the measuring distance does not have a significant impact on bandwidth aggregation. In the case of narrower bandwidth links to be aggregated to a pipe (like with L6’s situation), the flow control would affect the total throughput and utilization.