Authors:
Breno G. S. Costa
;
Marco Antonio Sousa Reis
;
Aletéia P. F. Araújo
and
Priscila Solis
Affiliation:
University of Brasilia, Brazil
Keyword(s):
Cloud Computing, Transient Server, Performance Analysis, Preemptive Machines.
Abstract:
A few years ago, Amazon Web Services introduced spot instances, transient servers that can be contracted
at a significant discount over regular price, but whose availability depends on cloud provider criteria and the
instance can be revoked at any time. Google Cloud Platform offers preemptive instances, transient servers that
have similar behavior and discount level to spot instances. Both providers advertise that their transient servers
have the same performance level as servers contracted on-demand. Even with the possibility of revocation
at the provider’s discretion, some applications can benefit from the low prices charged by these servers. But
the measured performance of both models, transient and on-demand, must be similar, and the applications
must survive occasional or mass server revoking. This work compares the performance and costs of transient
and on-demand servers from both providers. Results show there is no significant difference in performance
measured, but t
here is real cost advantage using transient servers. On Amazon Web Services a MapReduce
cluster composed of transient servers achieved a 68% discount when compared to the same cluster based on
on-demand servers. On Google Cloud Platform, the discount achieved was 26% but it can be bigger when the
clusters are larger.
(More)