Consistency Management in Cloud Storage Systems - Large Scale and Big Data: Processing and Management - page 348

the number of involved replicas necessary to maintain an acceptable stale reads rate
while allowing a better performance.

Algorithm 1: Harmony: Consistency Tuning Algorithm

Input : app_stale_rate is the predened consistency requirement.
Description: θ stale is estimated stale reads rate.
Output: X n is the number of replicas that retain the same consistency requirements.
if app_stale_rate ≥ θ stale then
Choose eventual consistency (Consistency Level = One)
else
Compute X n the number of always consistent replicas necessary to have
app_stale_rate ≥ θ stale
Choose consistency level based on X n

end

As shown in Algorithm 10.1, the default consistency level is the basic eventual
consistency that allows reading from only one replica. When such a level may not
satisfy the consistency requirements of an application due to the growing number of
stale reads, the number of replicas X n that should be involved in the reading requests
is computed as given in formula (10.3). All the following read requests will be per-
formed with consistency level X n .

− λ T

Ne (( 1 − )( 1

1

+ λλ ) − app _ stale ate _ λλ )

rP

X =

rw rw

(10.3)

n

− λ

r T

( − e )( 1 + λλ )

rw

10.6.4 h armony e valuation
To validate Harmony, a set of experimental evaluations was conducted on Amazon
Elastic Cloud Compute (EC2) [3]. For all experiments we ran Apache Cassandra-1.0.2
as an underlying storage system, and used the Yahoo Cloud Serving Benchmark!
(YCSB) [53]. We deployed Cassandra on 20 VMs on Amazon EC2. The goal of these
experiments is to evaluate system performance and measure the staleness rate. We
used Workload-A: a heavy read-update workload from YCSB! with two data sets
with the size of 23.85 GB for EC2 and a total of 5 million operations.

10.6.4.1 Throughput and Latency
We compare Harmony with two settings (two different tolerable stale read rates of
an application) with strong and eventual consistency (i.e., the ONE level). The first
tolerable stale read rate is 60% for Amazon EC2 (this rate tolerates more staleness
in the system implying lower consistency levels and thus less waiting time) and the
second tolerable stale read rate is 40% for Amazon EC2 (this rate is more restric-
tive than the first one, meaning that the number of read operations performed with