Workloads Performance & Optimizations
- 30 -
5.0 W
ORKLOADS
P
ERFORMANCE
& O
PTIMIZATIONS
5.1
Target Workloads for Ultrastar Memory
The Ultrastar memory drive can take advantage of one or more of the following
workload characteristics to achieve DRAM-like performance, which can be associated
with the sample relevant applications:
Table 5-1. Workloads and Applications
Probability-based memory
access pattern
[pre-fetch]
High concurrency access
pattern [asynchronous
memory load]
CPU intensive
[placement]
Predictable workloads:
Analytics, DBMS, etc.
Many processes, or highly
multi-threaded.
CAE, HPC
Memory access can be
predicted by:
Frequently similar access
pattern (sequential,
structured, etc.)
Application code analysis
Memory block associativity
Examples can be:
Databases tuples
In-memory columnar
database
Reading database indexes
Large in-core matrices
Block of memory stored from
a disk through DMA.
While relevant application
workloads can be row- or
column-store in-memory
databases used in analytics
workloads, such as SAP
®
HANA
®
, Oracle
®
, 12c, or
MySQL
TM
Concurrency can be due to:
Many threads
Throughput
(many independent jobs)
Multi-tenants
Relevant application
workloads can be:
Multi-tenant workloads
like container-based
virtual-shared web-
hosting server with
Docker
TM
, or
Virtualization-based
partitioning for example
with KVM.
Multi-threaded key-value
cache such as
memcached.
Distributed/shared data
grids and frameworks
such as Apache Spark
®
,
Apache Ignite
®
,
Aerospike, or Redis
TM
.
Those are workloads
which are heavy on
compute vs. memory
access. Relevant
applications and
workloads can be:
Multi-threaded linear
algebra workloads
with large matrices.
Parallel statistics
calculations on large
data.
