We now discuss the different workloads used in our study; these workloads have been used in the past to compare the performance of various concurrency control mechanisms [6][13]. LOWCON is intended to be a realistic low-contention workload; the others are intended to stress the system.
LOWCON.
This workload models a realistic system with low contention
(like those observed
in [27][16]).
Each client has a private region of 50 pages at each preferred server.
Each server has a _shared_ region of 1200 pages.
80% of a client's accesses go to its private region; the
rest go to the shared regions at its connected servers.
Thus, in this workload, a client can access 1250 pages at a preferred
server and 1200 pages at a non-preferred server (note that private
regions of other clients are not accessed by a client).
SKEWED.
As in LOWCON, each client has a private region of 50 pages at each
preferred server; together these regions consume 1000 pages. Each
server has a _shared_ region containing 250 pages. 80% of
a client's accesses go to its private region; 20% of the accesses go to
the rest of the database (or RDB) region consisting of all other pages at its
connected servers, including private regions of other clients.
This workload models
the case when a client has affinity for its own objects and is used for
checking contention with skewed sharing patterns.
HOTSPOT.
This is the same as SKEWED except there is a small region of
50 pages added to the database at each server (the shared region is thus 200
pages at each server). 10% of accesses go to the small region,
80% to the private region and 10% to the RDB. The small
region might be the top of a naming hierarchy that is accessed often by
clients; to capture the fact that such a region will not be modified
often, only one in every ten transactions modifies the small region.
This workload has the richest access patterns; it includes
both uniform sharing (all clients access the small region uniformly) and
skewed sharing (one client accesses its private region more frequently than
other clients). The workload is intended to stress the lazy
scheme because of high contention in the small region.
We could have added a small region to LOWCON, but in practice modifications to such a region are extremely rare and therefore its impact on lazy consistency would be insignificant.
HICON.
This is very unrealistic workload used in concurrency control studies to
model high contention. There are two
regions at each
server - a ``hot'' 250 page region that is accessed 80% of the time and a
1000 page ``cold'' region that is accessed 20% of the time.
Both hot and cold regions are shared uniformly
among clients.