Information Technology Reference
In-Depth Information
considering version management and personalized installations that are
sometimes required.
6.2.3 High-Performance Computing
In a general definition, high performance in a scientific context means
processing large data sets with large-scale resources. This imposes some
challenges for the design of a cloud computing solution for researchers,
including special attention to the details of the software and infrastructure
offered by cloud providers.
Even though there is an important variety regarding hardware and soft-
ware available, in general, cloud providers do not offer a platform designed
specifically for scientific computation. A platform like this would require
proper configuration of processing capability, high-throughput storage
devices, and operating systems with optimized libraries for making calculus.
Amazon Web Services (AWS) is actually making big efforts toward this by
offering its EC2 (Elastic Compute Cloud) Cluster Compute and Cluster GPU
instances. These instance types are specially designed for parallel applica-
tions that require a large amount of network communication. As their actual
offer, cluster instances can be configured with up to 244 GB of RAM mem-
ory, 10 Gbps of input/output (I/O) performance, 88 processing units, and
NVIDIA Tesla GPUs (graphics processing units) with “Fermi” architecture.
6.2.4 Data Communication
Although the available computing power is comparable to that found on
grid/cluster infrastructures, cloud providers still have a long road to face to
achieve the performance of these solutions. This seems to be especially true
when talking about communications, which according to Jackson et al. [2]
are the bottleneck for scientific cloud executions. Parallelization schemes often
require data sharing between processes executing on different machines.
Cloud infrastructure providers usually do not offer a dedicated data link or
any guarantees regarding network throughput. This means that a scientific
app running in a cloud has some limitations regarding the amount of data to
communicate while maintaining the required performance.
6.2.5 Costs
Scale economy is the biggest driver for cloud computing. The low costs at
which providers can acquire and maintain large data centers at geographi-
cally separate locations are the reason behind the success of the technology [3].
The idea of having access to thousands of servers just with a credit card and
with no initial acquisition costs is simply amazing.
Small- and medium-size research groups almost always work with a
small budget. An in-house infrastructure solution means that a great por-
tion of the money that was destined to buy investigation equipment and
Search WWH ::
Custom Search