Cisco Cisco UCS C250 M1 Extended-Memory Rack-Mount Server Libro blanco
48 © 2016 Cisco and/or its affiliates. All rights reserved.
Performance Characterization
The Cisco UCS Mini with Cisco UCS B200 M4 Blade Servers plus StorMagic software hyperconverged solution is designed to
handle branch-office and remote-office workloads that are more computation or network focused than I/O focused.
handle branch-office and remote-office workloads that are more computation or network focused than I/O focused.
This section describes the types of workloads most suitable for this solution and performance test results from a four-blade
configuration (starter kit) and a six-blade configuration (starter kit plus a two-blade expansion pack).
configuration (starter kit) and a six-blade configuration (starter kit plus a two-blade expansion pack).
Table 7 lists the workloads that are suitable to run on a Cisco UCS Mini and Cisco UCS B200 M4 plus StorMagic solution.
The application profiles and I/O patterns listed here do not represent a complete list of all workloads that can be made to run;
however, the majority of the workloads are expected to be similar.
The application profiles and I/O patterns listed here do not represent a complete list of all workloads that can be made to run;
however, the majority of the workloads are expected to be similar.
Table 7. Application I/O Profile: Cisco UCS Mini and Cisco UCS B200 M4 plus StorMagic Solution
Application Profile
Access Mode
Read:Write
Ratio
Block Size
Performance Metric
Web front end
Random
100:0
8 KB
IOPS and response time (milliseconds [ms])
Online transaction
processing (OLTP)
Random
80:20
8 KB
IOPS and response time (ms)
Decision support system,
business intelligence, and
video on demand (VoD)
Sequential
100:0 and 0:100
256/512 KB
Bandwidth or transfer rate (MBps)
Each disk provides approximately 140 to 150 IOPS with a response time of less than 20 milliseconds for 100 percent random-
read workloads when the virtual machine volumes are distributed across the entire disk. However, greater variations are noticed
in IOPS and response time when the virtual machine volumes are of smaller size, as a result of the mechanical head movement
between the inner and outer cylinders of the hard disk.
read workloads when the virtual machine volumes are distributed across the entire disk. However, greater variations are noticed
in IOPS and response time when the virtual machine volumes are of smaller size, as a result of the mechanical head movement
between the inner and outer cylinders of the hard disk.
For the purposes of this testing, 100 GB of hard-disk space was allotted to each virtual machine (which is typical in remote-office
and branch-office deployments, with four vCPUs, 6 GB of RAM, and 100 GB of storage on each mirrored volume.
and branch-office deployments, with four vCPUs, 6 GB of RAM, and 100 GB of storage on each mirrored volume.
Workload Scenarios
The following two scenarios were tested and validated for this solution:
• Scenario A, with one virtual machine per host (total of four virtual machines [starter kit] or six virtual machines [starter kit
plus 2-blade expansion pack]): This scenario sets the baseline for I/O performance, because no resource contention
occurs at the computing and network levels. The workload generated is purely I/O intensive and helps identify the maximum
achievable throughput and response time for the system.
occurs at the computing and network levels. The workload generated is purely I/O intensive and helps identify the maximum
achievable throughput and response time for the system.
• Scenario B, with three virtual machines per host (total of 12 virtual machines [starter kit] or 18 virtual machines [starter kit
plus 2-blade expansion pack]): This scenario helps estimate the achievable throughput and response time under load. It
uses three virtual machines per mirror to illustrate scalability.
uses three virtual machines per mirror to illustrate scalability.
In both scenarios, the access profiles listed in Table 7 were configured in the Iometer tool with a queue depth of 2. The tests
used an 8-KB block size because most applications use an 8-KB block size. A 4-KB block size was also tested, and the results
matched those for the 8-KB block size; hence, the graphs show the 8-KB block size.
used an 8-KB block size because most applications use an 8-KB block size. A 4-KB block size was also tested, and the results
matched those for the 8-KB block size; hence, the graphs show the 8-KB block size.