![](bg2.png)
determines a market price for the server (Agmon Ben-
Yehuda et al., 2013). If the market price increases and
stays above the bid, the server will be revoked, but
only after the two-minute notice period. While the SI
price remains below the bid, the SI remains available
and the customer will pay only the market price, even
if his bid is higher.
Figure 1 shows the market price variation, in a
one month period, of an M4.2xlarge type SI on the
us-east-1e zone. Each availability zone is a different
market and prices can be different. In the time frame,
the SI price reached the on-demand price (US$ 0.40)
less than 10 times and for short periods.
Due to the high probability of changing prices and
even the behavior of their transient servers in the short
term, Microsoft Azure Low-priority VMs will be ex-
cluded from this analysis, but they will be included as
future work as soon as the offer becomes stabilized.
1.1 Availability of Transient Servers
The availability of transient servers (in terms of aver-
age revocation time) can also vary significantly across
server configurations and on the basis of changing
market conditions. Unfortunately, cloud platforms do
not directly expose the availability statistics of tran-
sient servers, requiring users to infer them indirectly,
for example, through price history. Thus, it is chal-
lenging for a cloud system to select the most appro-
priate server configuration based on historical price
or availability data to meet its needs. Recent research
suggests that mitigating the risk of revocation requires
a parallelized system to diversify its resource needs
across various types of transient servers, further com-
plicating decision making (Sharma et al., 2016).
The problem is exacerbated by the large number
of transient server choices available from providers:
there are over 2500 SI options in AWS Elastic Cloud
Computing (EC2) and more than 300 GCP preemp-
tive instances. This is because each availability zone
has its own market value calculation for each avail-
able virtual machine configuration.
According to (Sharma et al., 2017), choosing a
server configuration based only on price can pro-
duce sub-optimal results. The authors cite an exam-
ple where server configurations with very low prices
can also see greater market demand and consequently
higher price volatility and more revocations. Frequent
revocations generate additional verification, check-
pointing, and system recovery efforts. Instead, they
suggest that choosing a slightly more expensive server
configuration and having a lower revocation rate can
produce lower overall costs.
Due to the challenges listed, cloud providers such
as AWS have begun offering server selection tools.
Amazon SpotFleet (AWS, 2015a) automatically re-
places revoked servers. However, SpotFleet has a
limited choice in terms of the combinations of server
configurations that it offers and does not solve some
of the challenges presented. Another tool, Amazon
Spot Bid Advisor (AWS, 2015b), can help users se-
lect servers based on price, but exposes only super-
ficial volatility information such as low, medium, or
high categorization.
An important consideration is that if transient
server performance was lower than the on-demand
server, and adding up the additional complexity of
dealing with revocations, the large discounts offered
by the providers would not be worthwhile. There-
fore, this study measures the performance of transient
servers using benchmarking software and compares
them to on-demand servers to verify whether the cost
decrease advertised by cloud providers is also accom-
panied by a performance decrease. In addition, a sce-
nario in which the use of a transient server is viable
will be implemented and the costs of execution in the
two server classes will be compared on both providers
that offer them.
The remainder of this article is divided into six
sections. Section 2 presents some related work. Sec-
tion 3 describes the experiments environment. Sec-
tion 4 describes the planning and the results of the ex-
periments. Section 5 performs an analysis of the per-
formance experiments. Section 6 presents a cost com-
parison of a MapReduce workload running on both
classes of servers, and in Section 7 conclusion and fu-
ture work are presented.
2 RELATED WORK
There are, in literature, several studies regarding spot
instances. Many of them try to predict SI prices and
find an optimal bid on the the spot market. The strate-
gies undertaken by these researches are diverse. Time
series forecasting is used by (Chhetri et al., 2017),
whose results, using three specific metrics, show that
successful estimation of bid prices in AWS spot mar-
kets is an implicit function of seasonal components
and extreme spikes in the spot price history. Another
study (Khandelwal et al., 2017) uses Regression Ran-
dom Forests (RRFs) to predict spot prices. The au-
thors use a one year trace of spot market prices and
compare the results achieved by RRFs with existing
non-parametric machine learning models. The paper
reveal that RRF-based forecast accuracy outperforms
other models.
In (Wolski and Brevik, 2016) a method is pro-
CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science
170