Browse Prior Art Database

Integrated Concurrency-Coherency Technique for I/O-Shipping or Data Sharing Supporting Record Locking

IP.com Disclosure Number: IPCOM000106709D
Original Publication Date: 1993-Dec-01
Included in the Prior Art Database: 2005-Mar-21
Document File: 4 page(s) / 156K

Publishing Venue

IBM

Related People

Bhide, A: AUTHOR

Abstract

I/O shipping is a promising method for obtaining scalable LAN servers because it is easy to implement. However, the naive form of I/O shipping involves sending too many messages. On a machine such as Risc System 6000* with high message overheads such an implementation will not perform well. We propose an integrated concurrency control and coherency technique which uses buffer invalidation (as against check-on-access) and which can support record-level locking.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 35% of the total text.

Integrated Concurrency-Coherency Technique for I/O-Shipping or Data Sharing Supporting Record Locking

      I/O shipping is a promising method for obtaining scalable LAN
servers because it is easy to implement.  However, the naive form of
I/O shipping involves sending too many messages.  On a machine such
as Risc System 6000* with high message overheads such an
implementation will not perform well.  We propose an integrated
concurrency control and coherency technique which uses buffer
invalidation (as against check-on-access) and which can support
record-level locking.

      In a partitioned database system with I/O shipping, the
database is partitioned among the nodes and data at remote nodes is
accessed by making I/O requests to the partition owning the data.  In
data sharing, all processors have direct access to all disks and thus
can run any transaction.  Logically, data sharing and I/O shipping
are equivalent; though the performance and availability
considerations are different.  A number of papers have proposed
methods for various concurrency and coherency protocols in such
architectures [1,3,4]  and studied their performance.  It has been
observed in [2,3]  that integrating concurrency and coherency
protocols can lead to considerable savings in terms of messages.
However, the protocols in [2,3]  can handle only page locking.
Moreover, [3]  uses a check-on-access coherency protocol combined
with primary copy locking.  They assume that a lock request message
needs to be sent for obtaining a lock even when a page is cached
locally and therefore it is possible to piggyback a "check-on-access"
message on it to ensure that the copy of the page is still valid.  We
believe that a buffer invalidation protocol combined with our
"dynamic lock authority"  assignment provides a superior method.

The following design considerations motivate this scheme:

1.  For OLTP workloads on relational databases, most accesses are
    through indices.  Index pages are read-mostly since they are
    updated very rarely.  Therefore, even workloads such as TPCA
    which update every base relation tuple they access involve a lot
    of read-only accesses to index pages.

2.  Thus, independent of other considerations (such as ability to
    piggy-back coherency messages on concurrency control ones),
    schemes which do "check-on-access"  are inferior to "buffer
    invalidation"  schemes sinc check-on-access schemes involve a
    message on every access (read or write).

3.  It is possible to design a "  dynamic locking"  scheme which
    eliminates messages for getting read locks under certain
    conditions.  Under these conditions (which occur frequently as we
    show later), such a scheme has no lock request message for the
    coherency protocol to piggy-back on for read accesses.

4.  Thus, buffer invalidation is a superior protocol for such a
    scheme.

5.  To avoid lock contention, it...