Cost-Benefit Analysis of Cloud Computing versus Desktop Grids page 4

GAMBAR

4.4 Completion times
The volatility and heterogeneity in VC systems makes
timely completion of task batches challenging. BOINC
has a number of mechanisms for ensuring time completion.
For example, project scientists can soft deadlines for tasks.
When the soft deadline of a task approaches, the local client
scheduler will increase the task’s priority relative to oth-
ers. In addition, the server-side scheduler uses the deadline
for determining timeouts, i.e., when another task instance
should be sent out.
With these mechanisms, task completion is usually done
at a high success rate. For example, in the World Commu-
nity Grid project (a non-profit project for volunteer com-
puting), 96.1% of tasks met their deadline out of 227,485
tasks [24].
Nevertheless, VC users should expect a stretch (defined
by the amount of time spent by the job in the system and
its execution time) of at least 5 according to our simula-
tion results in [21]. This is because the task deadlines are
usually high relative to the amount of actual work. The
median project deadlines are around 9 days, where as the
execution time per task is about 3.67 hours on a dedicated
3GHz host [6]. Recently, there has been promising results
in using predictive models for achieving fast turnaround
time [2, 19, 14]
By contrast, on EC2, platform construction takes a few
minutes to deploy an image. This assumes that the platform
is not overloaded. As resources are dedicated, application
deployment is instantaneous, and task execution and com-
pletion are relatively constant and low.

TABEL


5 Cloud Computing Costs
We present an overview of Amazon’s cloud services and
pricing [13] to be used in our calculations. Amazon has two
relevant cloud computing services. First, Amazon offers the
Elastic Computing Cloud service.EC2 charges each hour an
instance is running, and it offers instances with different
compute power and memory. The pricing for EC2 is shown
in Tables 1 and 2.
Second, in conjunction with EC2, Amazon offers the
Elastic Block Store (EBS) service. This provides reliable
and persistent storage with high IO performance. EBS
charges per GB of storage and per million IO transac-
tions. The pricing for EBS is shown in Table 3. Ama-
zon also offers the Simple Storage Service (S3). This
service provides access through web services to persistent
data stored in buckets (one-level of directories) along with
meta-data (key/value pairs). S3 charges per GB of stor-
age and HTTP requests concerning it. PersistentFS offers
a POSIX-compliant file system using S3 and is arguably
cheaper than EBS for mainly read-only data. However, for
volunteer computing projects, the cost difference between
S3/PersistentFS and EBS is not significant and does not
change our conclusions. Thus we assume all storage oc-
curs on EBS. We do not consider costs of snapshots, i.e.,
EBS volume backups to Amazon’s S3.

TABEL