Chapter
5 Implementing new MPSP algorithms
[pdf:336kB] [ps:754kB]
Many efficient SSSP and APSP algorithms and their implementations
have been proposed in the literature. However, only few of
them are targeted to solving the MPSP problems or problems
with fixed topology but changeable arc lengths or requested
OD pairs. These problems of course can be solved by repeated
SSSP algorithms. Other methods such as the LP-based reoptimization
algorithms (see Section 3.5.2) take advantage of previous
optimal shortest paths and perform either primal network simplex
methods (when arc lengths change) or dual network simplex
methods (when OD pairs change). More recent computational
experiments <cite>Bu93</cite> indicate these reoptimization
algorithms are still inferior to the repeated SSSP algorithms
which repeatedly solve shortest path trees for different requested
origins (or destinations, depending on which one has smaller
cardinality).
A practical implementation of Carré's algorithm <cite>Ca71</cite>
by Goto et al. <cite>GoOhYo76</cite> tries to
exploit the sparseness and topology of the networks. For networks
with fixed topology, their implementation first does a preprocessing
procedure to identify a good node pivoting order so that the
fill-ins in the LU decomposition phase are decreased. To avoid
unnecessary triple comparisons, they record all the nontrivial
triple comparisons in the LU decomposition, forward elimination
and backward substitution phases, and then generate an ad
hoc APSP code. Their method only stores O(m) data structures,
which is the same as other combinatorial SSSP algorithms but
is better than O(n²) as required in general algebraic
algorithms. However, this comes with the price of storing
the long codes of nontrivial triple comparisons, and may require
more total hardware storage. Even worse, the long codes may
not be compilable for some compilers.
In particular, we have tested a 1025-node, 6464-arc sparse
graph and generated a 500MB long code using the code generation
algorithm of Goto et al. <cite>GoOhYo76</cite>.
The code we generated could not be compiled even on a fast
Sun workstation with 1GB memory using gcc, a C compiler by
GNU, with optimization tags. If instead we only store the
arc index of all the triple comparisons, the code will be
short but still need temporary storage around 200MB to store
these indices. Therefore, code generation is not practical
for large networks.
In this chapter, we observe that Carré's algorithm
can be implemented combinatorially which resolves the need
of a huge storage quota by the code generation. Furthermore,
we can decompose and truncate Carré's algorithm according
to the indices of the distinct origins/destinations of the
requested OD pairs. Section 5.1 introduces some notation and
definitions appearing in this chapter. Section 5.2 describes
major procedures of our algorithm SLU, a sparse combinatorial
implementation of Carré's algorithm. Section 5.3 gives
detailed implementation issues, and techniques for speeding
up SLU. Implementations of algorithm DLU2, our proposed MPSP
algorithm that appeared in Section 4.3, are given in Section
5.4. Computational experiments including a sparse implementation
of the Floyd-Warshall algorithm, many state-of-the-art SSSP
codes written by Cherkassky et al. <cite>ChGoRa96</cite>,
and networks that we generate, are presented in Section 5.5.
Section 5.6 shows results of our computational experiments
and draws conclusions.
Summary
After these comprehensive
computational experiments, we have several observations.
- Among the eight different pivoting rules for sparsity,
the dynamic Markowitz and its variant that employs simple
tie-breaking technique usually produce fewer fill-ins
than other pivoting rules.
- Among the three different SLU implementations, the bucket
implementation SLU1 outperforms the other two implementations
that are based on binary heaps, for all cases. Similarly,
the bucket implementation of DLU2, DLU21, has better performance
than the heap implementation.
For general
larger networks, SLU1 and DLU21 can not compete with the
best of the state-of-the-art SSSP codes , except for acyclic
networks. Nevertheless, SLU1 and DLU21 do not perform
worst for all cases.
For dense graphs,
usually DLU21 performs better than SLU1; for sparse graphs,
usually SLU1 is better than DLU21. Unlike other SSSP algorithms
which perform consistently regardless the number of distinct
destinations, SLU1 and DLU21 perform relatively better
for MPSP problems of more distinct requested destinations
than for problems of fewer distinct requested destinations.
- The Asia-Pacific flight network is the only real-world
network tested in this thesis. Although it is sparse (112
nodes, 1038 arcs), DLU21 performs better than SLU1, which
in turn outperforms all of the implemented label-setting
codes. The label-correcting codes perform similarly to
SLU1 and DLU21. The Floyd-Warshall algorithm is in no
case competitive with other SSSP algorithms or our proposed
MPSP algorithms.
- In most SPGRID families, the label-correcting codes
usually perform the best, except for the SPGRID-PH family
for which the label-setting codes perform the best. SLU1
and DLU21 have better performance on the SPGRID-SQ and
SPGRID-WL families than other SPGRID families.
For each SPGRID
family, label-setting codes usually perform relatively
worse on smaller networks.
- SLU1 and DLU21 are usually slower than other label-correcting
and label-setting codes for most SPRAND families. Label-correcting
codes perform better for most sparse SPRAND families,
but label-setting codes perform better for most dense
SPRAND families.
When the range
of arc lengths decreases (e.g.,<=10 ), label-correcting
codes tend to perform much worse. When the range of arc
lengths increases on sparse graphs, label-setting codes
will perform only slightly worse.
- SLU1 and DLU21 are usually slower than other label-correcting
and label-setting codes for most NETGEN families. Label-correcting
codes perform better for most sparse NETGEN families,
but label-setting codes perform better for most dense
NETGEN families.
When the range
of arc lengths increases on sparse graphs, label-setting
codes tend to perform a little worse.
- SLU1 and GOR1 usually outperform other codes for all
SPACYC families except the SPACYC-PD family, for which
the label-setting codes perform asymptotically the best.
DLU21 performs asymptotically the worst for cases whose
arc lengths are all positive. However, all label-correcting
codes (except GOR1) and label-setting codes perform significantly
worse for cases with negative arc lengths.
Conclusion and Future Research
Although our
computational experiments are already extensive, more thorough
tests may still be required to draw more solid conclusions.
There are too many factors that may affect the performance
of MPSP algorithms, such as requested OD pairs, arc lengths,
network topologies, and node orderings.
In our experiments,
for each test case, we generate only one set of requested
OD pairs. Different requested OD pairs may affect the performance
since the distribution of requested pairs in the OD matrix
will not affect the SSSP algorithms but may affect our MPSP
algorithms.
By specifying
the same numbers of nodes and arcs but different random
seeds, we may generate several different test cases with
the same topology but different arc lengths. A more thorough
experiment could generate several such networks and test
all of the algorithms on these cases to compute their average
performance. Due to time constraints, in this chapter we
use only one random seed for each topology.
Different node
orderings will significantly affect our MPSP algorithms.
In this chapter, we choose a node ordering aimed at reducing
the total fill-ins in the LU factorization. As discussed
in Section 4.4, another ordering consideration is to group
the requested OD pairs and assign them higher indices. This
is difficult to test for randomly generated networks and
requested OD pairs. It may be worth trying for some real-world
applications where both the network topology and requested
OD pairs are fixed.
Our limited
results show that our MPSP algorithms seem to perform worse
for general larger networks. This may make sense since after
all SLU and DLU are algebraic algorithms, although we create
graphical implementations for testing. In particular, SLU
and DLU are more suitable for cases whose arc lengths are
randomly distributed for all cases since they are designed
to work independent of arc lengths. They are based on the
ordering of triple comparisons. The sequence of triple comparisons
is fixed and unavoidable in our MPSP algorithms. Thus, even
if there exists some arc with a very large length, our algorithms
could not avoid some trivial triple comparisons involving
this arc, although intuitively we know in advance these
operations are wasteful. Conventional SSSP algorithms, on
the other hand, are based more on the comparison of arc
lengths. For example, the greedy algorithm (Dijkstra's)
only does triple comparisons on nodes with the smallest
distance label in each iteration. Adding some "greedy"
ideas to our MPSP algorithms might help them avoid many
wasteful operations. However, it is still not clear how
to integrate these two different methodologies.
If we suspect
our MPSP algorithms may perform worse than conventional
SSSP algorithms for most cases, we can design an experiment
by generating the requested OD pairs most favorable to our
MPSP algorithms. In particular, since both SLU and DLU can
compute optimal distance faster for OD pairs with larger
indices, the most favorable settings for our MPSP algorithms
are: (a) find an optimal (or a very good) sparse node ordering,
and (b) in that ordering, generate k requested OD pairs
that span the k entries in the OD matrix in a line perpendicular
to the diagonal entries. That is, entries (n,n-k+1), (n-1,n-k+2),
…, (n-k+2,n-1), (n-k+1,n) will be the set of requested
OD pairs. Such a setting will force the conventional SSSP
to solve k SSSP problems, but will be most advantageous
for our MPSP algorithms since they are not only in the sparse
ordering but are also the closest to the "southeastern"
corner of the OD matrix in the new ordering. If experiments
on many random graphs with such settings still show that
our MPSP algorithms do not outperform the conventional SSSP
algorithms, we would conclude that our MPSP algorithms are
indeed less efficient.
The experiments
in this chapter are just a first step in evaluating algorithmic
efficiency when solving general MPSP problems. More thorough
experiments will be done in the future.
|