Database Reference
In-Depth Information
Fig. 6.
Performance improvement for dic-
tionary based DSM schemes for large col-
umn width
Fig. 7.
Performance comparison of differ-
ent DSM based schemes in ECOS for read
and write intensive workloadsf
also observed that evolving storage structures perform better than fixed storage
structures with minor performance gains. As we have discussed in Section 1, our
work is based on the ideology from Chaudhuri and Weikum presented in [6]. They
used the notion of “gain/pain ratio” to discuss the overall gain of their proposed
approach. They advocate the ideology of less complex, more predictable, and
self-tuning RISC-style components with minor compromise on performance to
achieve overall improvement in “gain/pain ratio”. Our results show the minor
performance gain, which should be a good achievement considering the overall
benefits we achieve in terms of simplicity, predictability, and self-tuning.
It can be observed in Figure 6 that dictionary based schemes performance
is improved and becomes comparable with standard 2-copy DSM scheme for
large tuple width. However, KDSM and MDSM still perform poor. We also
analyzed the performance difference for different DSM schemes on both the
read-intensive and write-intensive workloads. It is observed in Figure 7 that
for write-intensive workload DSM outperforms other schemes; however, for the
read-intensive workload differences in performance between the 2-copy DSM
and the dictionary based DSM schemes is minimum. This is a promising result
for dictionary based schemes, and it shows their potential to act as a better
alternative to 2-copy DSM after overcoming their short comings.
5 Related Work
Hierarchically-organized storage structures have already been in use in the data
warehousing domain. Morzy et al. in [14] proposed a hierarchical bitmap index
for indexing set-valued attributes. Later, Chmiel et al. in [9] extended that con-
cept to present hierarchically-organized bitmap indexes for indexing dimensional
data. Bender et al. proposed cache-oblivious B-Trees [5] that perform the optimal
search across different hierarchical memories with varying memory levels, cache
size, and cache line size. Fractal prefetching B+-Trees [8] proposed by Chen et
al. are the most relevant work for the ECOS and is similar in concept to cache-
oblivious B-Trees with an additional concept of prefetching. Fractal prefetching
