Professional Documents
Culture Documents
on
Megha Jain
IM-98-29
Abstract – Past research in message logging has focussed on studying the relative
overhead imposed by pessimistic, optimistic and causal protocols during failure free
executions. Applications face a complex trade-off when choosing a message logging
protocol for fault tolerance. In this empirical research, an evaluation of the performance
of these protocols during recovery is presented.
1. AIM
2. INTRODUCTION
3. LITERATURE REVIEW
On the basis of the literature review and from the understanding of the
concepts, I have formulated the following hypothesis: -
“There is a high correlation between the T[rec] and the T[rollfwd] in case of
receiver-based pessimistic, sender-based pessimistic and causal protocols;
and between T[rec] and T[replay] in case of optimistic protocols.”
And also,
“The best message logging protocol is one with the lowest T[rec].”
4.1 The Null Hypothesis
5. METHODOLOGY
5.1 Assumption
It is assumed that processes can fail by crashing, but the hardware is reliable.
Also, other overheads on a system are not considered.
For pessimistic and causal protocols, the recovery time (denoted by T[rec])
for a process consists of: -
1) T[chk], the time to restore the state of the failed process from its latest
checkpoint.
2) T[acq], the time to retrieve determinants and messages logged during
failure-free execution.
3) T[rollfwd], the time to roll-forward the execution of the process to its
precrashed state.
For all protocols, the value of T[rec] is determined by the three parameters: -
1) The number of concurrent failures, f. For optimistic protocols, multiple
failures may cause a process to rollback multiple times. For sender-based
pessimistic and causal protocols, multiple failures may complicate the task
of retrieving messages and determinants from other processes.
6. DATA COLLECTION
The data was taken from IEEE Transactions on Knowledge and Data
Engineering, Volume 2, No. 2, March/April 2000; from the research of
Sriram Rao, Lorenzo Alvisi and Harrick M. Vin.
7. FINDINGS
The mean is used as the measure of central tendency. The means of different
data sets: -
1) Receiver-based Pessimistic: Mean T[rec] = 168.39 sec.
2) Sender-based Pessimistic: Mean T[rec] = 187.50 sec.
3) Optimistic: Mean T[rec] = 176.60 sec.
4) Causal: Mean T[rec] = 197.90 sec.
1) Receiver-based Pessimistic: -
Table value of t = 2.179 (at 5% level of significance and 12 degree of
freedom). Calculated value of t = 0.10755
2) Sender-based Pessimistic: -
Table value of t = 2.179 (at 5% level of significance and 12 degree of
freedom). Calculated value of t = 0.02822
3) Optimistic: -
Table value of t = 2.201 (at 5% level of significance and 11 degree of
freedom). Calculated value of t = 0.0000038
4) Causal: -
Table value of t = 2.179 (at 5% level of significance and 12 degree of
freedom). Calculated value of t = 0.00267
8. RESULTS
Using the regression analysis, we can find the value of the dependent
variable, T[rec], given the values of the independent variables, T[chk],
T[acq], T[rollfwd], T[replay] and T[rollbck]. For this calculation, the values
of standard errors must be taken into account. The linear regression model
fits this calculation.
Using this feature of the descriptive statistical tools, the average value of the
cost of recovery (in terms of time in seconds), T[rec], was found out to be
minimum for Receiver-based message logging protocols (168.39 sec). This
suggests that (not considering other overheads) the best method is Receiver-
based message logging protocols.
8.4 Confidence Interval
The mean, standard deviation and the level of significance (here 5%) were
used to calculate the confidence interval. In our samples of data, we are 95%
confident that the mean is in the confidence interval i.e.: -
1) Receiver-based Pessimistic: 168.39 + 5.34 sec
2) Sender-based Pessimistic: 187.5 + 6.84 sec
3) Optimistic: 176.6 + 5.01 sec
4) Causal: 197.9 + 10.31 sec
It gives the upper and the lower limit within which the value can lie.
The t-test shows whether or not the hypothesis holds. If the calculated t-
value < table t-value, the hypothesis is accepted. Else, it is rejected and the
alternate hypothesis is accepted.
1) Receiver-based Pessimistic: -
0.10755 < 2.179, the hypothesis “there is a significant dependence of
T[rec] on T[rollfwd]”; is accepted.
2) Sender-based Pessimistic: -
0.02822 < 2.179, the hypothesis “there is a significant dependence of
T[rec] on T[rollfwd]”; is accepted.
3) Optimistic: -
0.0000038 < 2.201, the hypothesis “there is a significant dependence of
T[rec] on T[replay]”; is accepted.
4) Causal: -
0.00267 < 2.179, the hypothesis “there is a significant dependence of
T[rec] on T[rollfwd]”; is accepted.
8.6 Significance of R-square test
The closer the value of R-square is to 100%, the more is the probability that
we have included the most important variables in our analysis.
It tests the significance of R-square test and the overall validity of the
model. If calculated F-value <= table F-value, the null hypothesis is
accepted. Else, it is rejected. For the model to be acceptable, the null
hypothesis should be rejected.
1) Receiver-based Pessimistic: -
0.62777 < 4.18; the null hypothesis is rejected.
2) Sender-based Pessimistic: -
0.91958 < 4.18; the null hypothesis is rejected.
3) Optimistic: -
0.48144 < 4.10; the null hypothesis is rejected.
4) Causal: -
0.20094 < 4.18; the null hypothesis is rejected.
Thus, the null hypothesis is rejected and the overall model is held
significant.
9. CONCLUSION