Optimizing Infrastructure for High Value Compute Environments

Premise

High value computing has unique compute, network and storage infrastructure requirements. The key systems requirements are low or very low latency, high bandwidth and high availability designed to lower software costs, improve time-to-value and help maintain high levels of data integrity and system availability.

Executive Summary

Wikibon researched the differences between performance optimized and cost optimized systems for a sample high-value compute environments, and concluded that the total 3-year cost of cost-optimized configurations and software were 63% high than performance optimized configuration. This is illustrated in Figure 1 below.

The key differences are higher costs for performance optimized hardware, but lower costs for the software required in high value compute application, as well as lower operational costs.

Definition of High Value Compute

Wikibon defines high value compute as:

1. Applications where the software costs significantly exceed the hardware costs. These software costs typically include high function databases such as Oracle, IBM’s DB2, or Microsoft SQL Server, and may include other performance critical applications suites such as SAS (Statistical Analysis System) from by the SAS Institute which provide mission critical advanced business intelligence and predictive analytics.
2. These tier 1 mission critical applications are required to perform at the highest possible levels, and require low-latency high bandwidth computing, which allows business application value to be optimized.
3. These tier 1 mission critical applications require a complete and fire-tested set of very high availability functionality, remote replication functionality allowing large-scale consistent replication and recovery to and from multiple locations;
4. These tier 1 mission critical applications are supported by a set of well-established performance and availability services that:
   1. Deeply understands the compute and software technologies;
   2. Understand how to integrate and project manage those technologies to meet the specific organization and industry application imperatives;
   3. Can prove those requirements have been met.

Figure 2 puts different workloads into a compute value spectrum. Low and to the left is low value compute, where infrastructure is optimized for cost. High Value Compute is up and to the right, where the software and application requirements require low latency servers, low latency or ultra low-latency storage and high-bandwidth point-to-point network infrastructure. The infrastructure requirements for high value applications are in stark contrast to lower-value applications (such as simple scale-out web-based applications) where the cost of hardware compute is often the dominant business line-item.

Optimizing Compute Requirements for High Value Computing

Optimizing the compute requirements for high value computing revolve around a core set of principles:

1. Reducing the amount of software that is required to be licensed, especially software that is charged on a per-core bases;
2. Providing optimum time-to-value for business, either as very fast response time or low elapsed time-to-value;
3. Ensuring that aggressive system level agreements (SLAs) for system availability, system recovery (RTO^[1]) and data integrity (RPO^[2]) can be met.

High Value Compute: Case Study Assumptions

The easiest way to illustrate how to optimize environments is to take a case study and show the difference between the an infrastructure system optimized for compute cost, and an infrastructure system optimized for application value when applied to high value compute applications:

System Optimized for Compute Costs:

• Server: Optimized for cost per core;
• Storage: Tier-2 traditional all-flash storage array with 1ms average IO, medium bandwidth and data reduction using compression & de-duplication;
• Network: Ethernet;

System Optimized for High Value Compute:

• The applications all used high-value application middleware. The reference architecture is that 80% of the available server cores were applied to applications and system administration, 6% to Oracle Database, 10% to Microsoft SQL Server, and 4 % to SAS.
• Server: Optimized for speed per core (large faster main storage, fewer but faster cores)
• Storage: Tier 1 high performance all-flash storage array with .35ms or lower average IO and high bandwidth with no compression de-duplication. The reference architecture for the Tier 1 performance storage is an EMC VMAX all-flash array, which includes low latency writes, a large global cache for reads, high overall bandwidth, InfiniBand connected storage controllers, and strong integration into Oracle and Microsoft database management systems. The backup and recovery systems were EMC equipment and software.
• Network: High bandwidth low latency network between systems. The reference architecture is point-to-point InfiniBand/Fibre Channel.
• Data availability, integrity, backup and recovery was common across all applications, and based on  EMC deduplication Data Domain appliances.
• Management of Oracle applications was through plug-ins into Oracle Enterprise Manager (OEM), and though plug-ins into Microsoft System Manager.

Key Assumptions:

• Multiple high value compute applications, including mission critical applications using Oracle DB, business critical applications using Microsoft SQL Server and real-time business analytics using SAS;
• Aggressive RTO and RPO SLAs;
• Need for consistent backup and recovery mechanisms;
• Tier 1 storage environment;
• 20% of cores used for Oracle DB, SQL Server or SAS;
• 80% of cores used for application and system administration;
• Cost of Oracle DB software is $35,000 per core after discount (6% of cores);
• Cost of Microsoft SQL Server is $20,000 per core after discount (10% of cores;
• Cost of SAS is $22,000 per core after discount (4% of cores);
• Cost of DBA or Equivalent is $200 per Hour;
• Cost of System Admin or Equivalent is $140 per Hour;

High Value Compute: Case Study Results

Figure 2 below shows the results of the analysis. The key difference is the reduction in software licensing costs from $5.2 million to $2.5 million over three years. 42% of this software cost was Oracle software, 40% Microsoft SQL Server, and 18% SAS. This reduction was because the high performance environment needed far fewer cores.

The detailed calculations in Figure 3 are shown in Table 1 below. The total number of cores in the compute environment is 440 for the cost optimized environment, and 200 for the performance optimized environment. The cost per core of the performance-optimized hardware is $6,664, over four (4) times more expensive than the $1,506/core for the cost-optimized hardware.

The cost for DBAs and system administration is significantly lower in the performance-optimized systems, as less tuning and problem determination is required.

Table 2 below shows the business case for performance-optimized systems. The net present value of the benefit of using performance optimization for high value compute is $2.5 million over three (3) years. As the overall capital costs are also lower, the breakeven is 0 months.

Conclusions & Recommendations

High value compute environments are best served by high performance servers with significant amounts of very fast DRAM, fast cycles times and fewer cores, together with very low-latency/high bandwidth tier1 storage arrays. The networks in these high performance systems are recommended to be high bandwidth/low-latency, and point-to-point.

Action Item

Wikibon strongly recommends that performance optimized compute models be included in RFPs for high value compute applications and environments.

Footnotes:

Table 3 below gives detailed assumptions that are used in the calculations in Tables 1 and Table 2 above.

Note^[1] Recovery Time Objective – Time to fully recover after application failure or disaster.

Note^[2] Recovery Point Objective – Amount of data that could be lost after application failure or disaster.