IBM Systems Magazine, Mainframe - November/December 2016 - L22

TRENDS

IDEAL PLATFORM
Co-locating analytics on Linux on z Systems adds value

inux* on z Systems* is well regarded as an ideal consolidation platform for
database workloads. Databases traditionally have been hosted on bare metal
servers, and with proliferation of commodity x86 servers, it's not uncommon to
find hundreds, if not thousands, of distributed servers running database workloads
in a data center. This has resulted in complexity with explosive software license
charges and significant operational costs for administration, floor space and energy.
Avijit Chatterjee
is an analytics
evangelist in the
IBM Competitive
Projects Office.

Some companies have
addressed their server proliferation challenges by sweeping their
data center floor of distributed
servers and consolidating them
on Linux on z Systems. Laboratory studies conducted by the
IBM Competitive Project Office
have shown that this exercise
often cuts total cost of ownership
in half because of tremendous
software license savings achieved
from drastic reduction in cores.
While these databases could be
running either online transaction
processing (OLTP) or analytical
workloads, more often than not,
they are systems of record (SoRs)
running OLTP; the analytical systems are special-purpose appliances in most customer shops.
This article presents two studies
that demonstrate tremendous
value in co-locating the analytics
workload with the systems of
record on Linux on z Systems.

An Important Role

Analytics drives competitive
differentiation for many
companies. Therefore, it is hot
on CEOs' radar. As a result,
new chief data officer and chief

analytics officer positions have
been added in the C-suite of
most companies.
Digital business is all about
speed, which drives the need
for real-time analytics directly
on top of the OLTP data. This
type of analytics is also called
operational analytics. There's
also deep analytics that
consists of queries run against
historical and unified data from
an enterprise stored in data
warehouses and marts.
Operational analytics consist of short running queries,
whereas deep analytics are long
running, for they often involve a
full table scan of extremely large
fact tables consisting of billions
of rows of measures that are then
sliced and diced using a dimension table consisting of attributes
such as product, demographics,
time, etc.
We conducted two studies
using best-of-breed in-memory
technologies Spark and DB2*
BLU for the operational and deep
analytics scenarios, respectively.
The first study involves
an operational analytics
scenario, where a supervisor

of a brokerage firm wants to
examine the performance of
brokers in terms of total trade
amount handled in real time.
The brokerage OLTP database
in this case is hosted on Linux
on z Systems. Spark SQL was
used to run the aggregate query
to measure the performance
of the brokers. The database
is over 100 GB in size, and the
trade table across which the
aggregate query is computed has
over 360 million rows of trading
transactions.
We ran an apples-to-apples
test running Spark on the same
number of cores on z Systems as
an x86 server and obtained up to
3x higher throughput co-locating
Spark with the OLTP database
on the same platform. Spark
provides a universal OS for endto-end analytics and, though we
tested with SparkSQL, we expect
similar throughput benefits
for co-location for other Spark
modules, such as MLib, GraphX
and Streaming.
The second study is for deep
analytics on a 2.5 TB data mart,
containing data from a retail
store with multiple channels (in

22 // NOVEMBER 2016 Linux on z Systems

Linux_on_z_1116.indd 22

9/29/16 1:15 PM

Table of Contents for the Digital Edition of IBM Systems Magazine, Mainframe - November/December 2016