### Table 3: Pre-copy checkpoint overhead

"... In PAGE 6: ... We use pre-copy to avoid blocking the processes while the checkpoint is written on the stable storage. Table3 presents the running time with pre-copy checkpointing. With full checkpointing, the performance degradation is dependent of the amount of data to be saved, due to the latency in accessing file servers.... ..."

### Table 11: M ops for IK forms in PA-Risc without pre-copy

1994

"... In PAGE 9: ...Table11 ) when no precopy is performed. The catastrophic e ect is only produced for block section size 153, since this is larger than the number of TLB en- tries.... ..."

Cited by 29

### Table 20: Comparable time bounds for alternative 1, data copying methods

"... In PAGE 50: ... The measured values in Table 17 are used to calculate the probable transmission time. In Table20 , these calculated values are shown. The values in the table are calculated applying the simplified model in Equation 2.... In PAGE 50: ...Table 20: Comparable time bounds for alternative 1, data copying methods Please notice that the values in Table20 are not calculated with consideration to the requirement of the switching capacity. Those values are only used to compare the methods for solving this sub-domain.... In PAGE 51: ... If alternative 1, namely copying the data directly from the slave memory, is used it will take 1.76 seconds to perform the data copying part of the program, see Table20 . If the slave memory is copied via the VME memory instead, it will take 1.... ..."

### Table A.1: Performance of the Minepi algorithm for serial (a) and parallel (b) episodes with repetition from data set 4; maximum time bound win = 60 s. ( = virtual memory exceeded.

1999

Cited by 16

### Table 3: Comparison of solution quality of search with and without abstraction. Time-bound

1992

"... In PAGE 14: ...sed. For the search without abstraction, the latter vector is the one used as constraints. Recall that by this vector the constraints C n and C d are considered more important than the constraint C l . Table3 describes the comparisons in the quality of solutions found, taking both sets of constraints as soft ones, and giving both problem-solvers an equal time bound (18 CPU seconds on a Sun4 Sparc Station). Each item in the table is a vector (x;; y), where x is the number of remaining violations with the most important constraint, and y the least important one.... ..."

Cited by 4

### Table 2: A comparison of PRE and copy propagation algorithms w.r.t, redundant loads eliminated.

"... In PAGE 7: ... To the best of our knowledge, this paper is the rst demonstration of these optimisations in Java programs by considering the interprocedural modi cation side effects in the presence of dynamic class loading. Table2 compares the two PRE algorithms and the two copy propagation algorithms in terms of the number of redundant loads eliminated. We see convincingly that our... ..."

### Table 1). The dotted lines give the average gains without the use of problem sizes, and the solid lines are for the gains with the regression. The graphs show that the use of sizes usually, though not always, leads to a small improvement. The apparent advantage of the regression in delay apos;s learning is mostly due to the choice of low time bounds for problems 9 and 10, which cannot be solved in feasible time. This luck in setting low bounds for two hard problems is not statistically signi cant. If the algorithm does not use problem sizes, it hits the time bounds of 16.9 and 14.0 on these problems (see Figure 5) and falls behind in its per-problem gain.

1997

"... In PAGE 3: ... The application of a method to a problem gives one of three outcomes: it may solve the problem; it may terminate with failure, after exhausting the available search space without nding a solution; or we may interrupt it, if it reaches some pre-set time bound without termination. In Table1 , we give the results of solving thirty transportation problems, by each of the three methods; we denote successes by s, failures by f, and hitting the time bound by b.... In PAGE 4: ...1 s 5.4 f 4 Table1 : Performance of apply, delay, and alpine on thirty transportation problems. Note that these data are only for illustrating the selection problem, and not for a general comparison of these search techniques; their relative performance may be very di erent in other domains.... In PAGE 4: ... Also note that the selection technique does not rely on speci c properties of prodigy; it is equally applicable to selection among multiple methods in any AI system. Although each method outperforms the others on at least one problem (see Table1 ), a glance at the data reveals that apply apos;s performance in this domain is probably the best among the three. We use statistical analysis to con rm this intuitive conclusion, and show how to choose a time bound for the selected method.... In PAGE 12: ...pply apos;s estimate of the maximal-gain bound, after solving all problems, is 9.6. It di ers from the 11.6 bound, found from Table1 , because the use of bounds that ensure a near-maximal gain has prevented su cient exploration. delay apos;s total gain is 115.... In PAGE 12: ...elay apos;s total gain is 115.7, or 3.9 per problem. If we used the data in Table1 to nd the optimal bound, which is 6.2, and solved all problems with this bound, we would earn 5.... In PAGE 12: ...3 per problem. The estimate based on Table1 gives the bound 11.0, which would result in earning 12.... In PAGE 14: ...0. In this experi- ment, we rst use the thirty problems from Table1 and then sixty additional transportation problems. The horizontal axis shows the number of a problem, and the vertical axis is the running time; we mark successes by circles and failures by pluses.... In PAGE 19: ... We denote the number of sample problems by n, the problem sizes by size1; :::; sizen, and the corresponding running times by time1; :::; timen. In Figure 12, we give the results of regressing the success times for the transportation problems from Table1 . The top three graphs show the polynomial dependency, whereas the bottom graphs are for the exponential dependency.... In PAGE 20: ... We also allow the user to set a regression slope, which is useful when the past data are not su cient for an accurate estimate. If the user speci es a slope, the system uses her value in the regression; however, it compares the user apos;s value with the regression estimate of Table1 1, determines the statistical signi cance of the di erence, and gives a warning if the user apos;s estimate is o with high probability. Note that the least-square regression and related t-test make quite strong assumptions... ..."

Cited by 2

### Table 22: Comparable time bounds for alternative 2, data copying methods

"... In PAGE 52: ... In order to be able to compare this alternative to the first alternative, we use the measured values in Table 17 to calculate the probable transmission time. Table22 shows these calculated values. Please notice that the values in this table are not calculated with consideration to the requirement of the switching capacity.... ..."

### Table 1. Time bounds for different solutions to the Time Queue problem

"... In PAGE 3: ... The standard heap described by Williams [18] can be modified to use fingers by adding a dictionary that stores the position in the heap for each element. The heap solution (heap in Table1 ) even works if the maximum duration is unbounded and it only needs O(N) space. The model used is the pointer machine model [11].... In PAGE 3: ...) [13,15]. However, the stratified tree needs O(C + N) space. The model is the pointer machine model. Willard shows how perfect hashing (see [6, 8]) can be used to improve the space bound to O(N) for the stratified tree [17] (vEB-W in Table1 ). The model is the RAM model [14] of the stronger cell probe model [19] due to the hashing.... In PAGE 3: ... They use O(C + N) space in the Yggdrasil implementation [2] of the RAMBO model [9]. So far we have seen the bounds in Table1 , with the Calendar queue (CQ)... ..."