Following Performance. Leading to Prosperity. TM

White Paper

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions April 2016

Sequor Systems Toll Free: 1 800 932 4387 www.sequorsystems.com E-Mail: [email protected]

Contents 1 Abstract .......................................................................................................... 1 2 The Data Center Landscape ......................................................................... 2 2.1 2.2 2.3 2.4

Background.................................................................................................... 2 Data Center Growth Potential ........................................................................ 3 Scaling to meet Demand ................................................................................ 4 Disruptive Technologies in Computer Software.............................................. 4

3 Business Challenges .................................................................................... 5 3.1 3.2

Cost and ROI Concerns ................................................................................. 5 Web and Application Server Software ............................................................ 6

4 Investigation of Web Server Software Efficiency ....................................... 8 4.1 4.2 4.3

Objective ........................................................................................................ 8 Overview ........................................................................................................ 8 Performance and Efficiency Testing ............................................................... 8

5 Results ......................................................................................................... 10 5.1 5.2 5.3 5.4

HTTP Results .............................................................................................. 11 HTTPS Results ............................................................................................ 11 Performance ................................................................................................ 13 Efficiency ..................................................................................................... 14

6 Estimated Capital and Operating Cost Savings........................................ 15 6.1 6.2 6.3

Capital Cost Analysis ................................................................................... 16 Operating Costs Analysis ............................................................................. 17 Total Cost of Ownership............................................................................... 18

7 Conclusions ................................................................................................. 19 8 Appendices .................................................................................................. 19 8.1 8.2 8.3 8.4

Test Configuration Hardware Specifications................................................. 19 Test Configuration Software Specifications .................................................. 20 About the Authors ........................................................................................ 21 Further Reading ........................................................................................... 21

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

3

1 Abstract Data centers are facilities that provide the infrastructure that powers the internet, the cloud, and the enterprise. Substantial capital and operating costs are associated with the construction, management, and operation of data center facilities. This white paper shows how technology innovations developed by Sequor Systems may cut capital and operating costs by a factor of approximately 2 to 4 times or greater. Due to growth of the internet and services delivered via “the cloud”, data centers are expanding their capacity and services to meet demand. This is causing increased pressure on existing infrastructure and requires increased capital and operating expenditures. For the better part of the last decade, the industry has been searching for solutions. According to Gartner1: Perhaps the most compelling incentive currently available for organisations to ‘think and act green’ is the potential to reduce costs, hence the sharp focus on lowering energy consumption. “Organisations are increasingly deploying more computing power,” said Rakesh Kumar, research vice president at Gartner. “These systems require considerably more power and cooling than the last generation of hardware. Because global energy prices are rising, there is a significant increase in data center operational budgets. Most large enterprise IT organisations typically spend in the region of four to eight percent, in some cases 10 percent, of their total IT budgets on energy. But the twin factors of power hungry hardware and rising energy costs could lead to this figure rising by up to four times within five years. This will put pressure on the CIO to act, and is placing power consumption high up the IT agenda. Source: Gartner Newsroom, November 7, 2006

Central to these “green” strategies is the management of energy consumption. Increasing energy costs, as well as energy market volatility complicate the service provider’s ability to plan and scale to meet industry demand. In order to better manage operational costs related to electricity consumption, data center operators have adopted metrics and strategies related to Power Usage Effectiveness2 (PUE). In order to further improve the efficiency of server equipment, Tier44 Technologies has created the PAR43 Server Energy Efficiency Measurement. The PAR4 methodology rates server hardware according to its energy efficiency score. Correct determination of PUE is potentially difficult and error-prone. According to Mike Jansma, Co-founder and Chief Marketing Officer of Enlogic, “Data centre managers are being blindly driven by the belief that PUE is the magic metric. It’s not. It’s useful, but only when used in combination with other factors to measure performance.”

1

http://www.gartner.com/newsroom/id/498224

2

http://en.wikipedia.org/wiki/Power_usage_effectiveness

3

http://par4.org/

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

1

Efforts4 made to measure and improve efficiencies include metering of electricity consumption, optimizing HVAC, modernizing equipment, and use of virtualization. Although these approaches have made improvements in PUE, perhaps the greatest enabler of high efficiency to reduce capital and operating costs is disruptive software technologies.

2 The Data Center Landscape 2.1

Background

Published in 2013 by Digital Realty Trust, a survey of data center operations of large North American companies (having either $1 billion or more in revenues or 5,000 or more employees) shows that:  The average data center facility is over 15,000 square feet;  66% of respondents have built or acquired a new data center in the past two years;  The average data center power load is 2.6 mega watts and the trend in power use is increasing; and  81% of data center operations meter power use. In view of costs expended in the development and maintenance of service capacity and operating margins, data centers are seeking to identify and implement strategies to improve operational efficiency and to reduce capital and operating expenses. According to an article published in 2010 by Gartner5, “Energy-Related Costs Account for Approximately 12 Percent of Overall Data Center Expenditures.” A 2013 report published by Digital Realty6 indicates that power usage is continuing to trend upwards in the data center:

4

https://www.facebook.com/notes/facebook-engineering/building-efficient-data-centers-with-theopen-compute-project/10150144039563920 5

http://www.gartner.com/newsroom/id/1442113

6

http://investor.digitalrealty.com/Mobile/file.aspx?IID=4094311&FID=16650607

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

2

Source: Digital Realty North America Campos Survey Results, January 2013

Another aspect of this key trend, and one directly impacting data center management, is the cost and availability of energy. According to Gartner7, energy and related expenditures represent the fastest-growing data center cost, constituting fully 12% of total data center expenditures as of 2010. Although energy prices see periods of volatility, initiatives and incentives to “go green”8 reinforce the need to factor energy consumption and efficiency into data center operations and capacity planning.

2.2

Data Center Growth Potential

According to a joint report by Gartner, Bank of America, and Merrill Lynch, the enterprise and cloud computing marketplaces are expected to triple in worth to $31 billion by 2016. Correspondingly, analysts at TechNavio project the Global Hosted Virtual Desktop market to grow at a CAGR of 63.7 percent over the period 2013-2018. Both of these market segments rely on efficiency, availability, and capacity in data centers. In order to meet growing market need, data centers are being pressured to expand capacity. According to Digital Realty, 81% of those enterprise data center respondents surveyed expect server density to increase.

7

http://www.gartner.com/newsroom/id/1442113

8

http://www.greendatacenternews.org/articles/share/785030/

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

3

2.3

Scaling to meet Demand

The data center has thrived in part due to continuing advancements in computing performance and efficiency. Historically, increased computing efficiency and performance have followed Moore’s law9, thus satisfying the increased computing demands placed upon data centers. Until the last decade, Moore’s Law anticipated a doubling of computer performance roughly every 18 months; however this rate of growth has been stymied by physical limitations in CPU clock speed. To facilitate continued performance improvements, parallelization has become a central aspect of modern CPU architecture. In response to processor frequency limitations, manufacturers have developed multicore processors, which allow computers to process separable tasks concurrently. Software infrastructure, however, has not yet been adapted to the parallel processing paradigm, nor has it addressed the scalability challenges of multi-core computing. Without such innovations in computer software, the potential of multi-core computer systems will remain unrealized. Failure to effectively leverage these advances in processor technology constitutes a serious impediment to business growth and a decrease in ROI. Non-blocking algorithms provide dramatic performance and efficiency improvements in software, often greater than factors of 5 times. These improvements translate directly to improvements in IT load required to meet capacity demands. This technology promises to deliver substantial benefits that support business growth and increase ROI.

2.4

Disruptive Technologies in Computer Software

Disruptive technologies in the software marketplace provide important opportunities for meeting increased customer demands, while simultaneously achieving capital and operational cost goals. One such technology that demonstrates strong performance gains is “non-blocking synchronization”, also known as “lock-free” programming. This breakthrough technique allows concurrent access to shared data without the delays otherwise incurred as exclusive access is obtained, usually on the part of the kernel. Thus, as program execution continues essentially uninterrupted, CPU resources are more effectively utilized and not wasted. The field of lock-free programming is emerging as a viable means to improve scalability and efficiency in multi-processing and parallel processing environments. According to Dr. Damian Dechev, a leading subject matter expert in the field, "...Studies show that non-blocking synchronization can deliver significant performance improvements."

9

http://en.wikipedia.org/wiki/Moore%27s_law

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

4

3 Business Challenges Data Center trends are strongly influenced by the following marketplace shifts:    

Demand for more services offered online (via the “cloud”); Increased processing demand by new methodologies including “big data”; Demand for increased efficiency in energy usage (“green” computing); and Increased use of virtualization and hosted virtual desktops.

According to research by Digital Realty Trust, “3 in 5 [respondents] say that implementing an internal cloud is an extremely important reason for expansion” and “larger companies ($5B+ revenues) are expanding in fewer but larger data centers with more available power.”

3.1

Cost and ROI Concerns

As data centers are expanding services in areas including the cloud and virtualization, strategies to lower costs, reduce risks, and improve ROI are being evaluated. Consolidation of computing resources and management of PUE has direct affect on ROI in both capital costs and operational costs. 3.1.1

Total Cost of Ownership

According to Bruce MacDougall, President of Internetworking Atlantic, Inc., “… capital costs have a direct dependency on the performance of the underlying IT and operational costs have a direct dependency on the efficiency of the underlying IT… both factors are critical to the bottom line.” 3.1.2

Capital Costs

Data center capital costs are based upon factors such as total building space, IT capacity, power capacity, cooling system, air distribution, backup power generation, and level of redundancy required. Capital costs are a function of capacity of the IT load. For a given IT load, the number of servers required is dependent upon their performance. Server performance is dependent upon software performance. Higher performance translates into reduced requirements of building space, subsystems and redundant capacity. Higher performing IT may have a dramatic and positive impact on capital costs.

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

5

3.1.3

Operational Costs

Data center operating costs are based upon factors including the supply and maintenance activities involved in delivering IT services. Operating costs are a function of PUE of the IT load. Server electricity usage is directly dependent upon the efficiency of the hardware and software as well as the power utilization of data center subsystems and infrastructure. More efficient IT may have a dramatic and positive impact on operating costs.

3.2

Web and Application Server Software

Web and Application server software provides essential services across a network. These services act as the "gateway" through which data flows in and out of a data center and are a key factor in their efficient operation. Demand for greater computing performance, efficiency, and stability is reshaping the landscape of Web and Application software and services. Within the broader IT marketplace, such forces translate into a competitive need to adopt faster, greener, and more scalable solutions. The Web Hosting and Server market is keenly sensitive to these industry trends and how they impact current and projected industry demand. 3.2.1

Market

Currently, the three web and application server market leaders are Apache, Microsoft, and NGINX (pronounced “engine X”). This marketplace has, historically, been largely dominated by the Apache web server. Apache is open source and free to download and use; both individuals and businesses benefit from the free availability of Apache and its minimal capital costs. Operational costs are, however, clearly a growing concern. 3.2.2

Trends

Although Apache has been a dominant vendor in the web server marketplace, a recent survey shows enterprises and web-based businesses shifting to alternatives. Of the million busiest web sites, NGINX now claims a market share exceeding 40%10. This shift is being driven by both capacity and availability concerns where the cost of poor software performance and stability overwhelm the savings of otherwise free software. According to a 2007 paper11 presented by Bryan Veal and Annie Foong to the ACM/IEEE Symposium on Architectures for Networking and Communications Systems on behalf of Intel, web servers in general, and Apache 2 in particular, scale poorly on multi-core hardware.

10

http://news.netcraft.com/archives/2014/10/10/october-2014-web-server-survey.html

11

http://www.cse.wustl.edu/ANCS/slides/Bryan%20Veal%20ANCS%20Presentation.pdf

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

6

As a central component of internet infrastructure, the adoption of web server platforms other than Apache illustrates the growing demand for a high availability, high performance software solution in the web server marketplace. In riding this wave of innovation, data centers have been searching for solutions that will provide better performance and efficiency than the Apache web server, as volatility in the web server marketplace demonstrates:

Source: Netcraft October 2014 Web Server Survey

According to Netcraft,12 “Apache leads in this market with a 47.5% share, and Microsoft also performs well with 30.7%, but both have been gradually falling over the past few years as a result of NGINX strong growth. NGINX gained more than 17,000 additional web-facing computers this month, helping to bring its market share up to 10.3%.” The clear implication of this shift is a renewed industry focus on achieving growth and cost savings through server performance, efficiency, and stability.

12

http://news.netcraft.com/archives/2014/10/10/october-2014-web-server-survey.html

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

7

4 Investigation of Web Server Software Efficiency 4.1

Objective

The objective of the investigation is to determine the baseline performance and efficiency figures of those web and application server software solutions currently used within data centers as compared with the Sequor web and application server. In order to determine minimum performance and efficiency improvements expected in a conventional data center environment, a server configured with a lower-density, dualcore (four hyper-threads) CPU package was utilized. The Sequor web server uses the innovative non-blocking synchronization technology described earlier. This and other techniques developed and employed by Sequor have been shown to scale even more effectively on higher density CPU packages, therefore use of a lower density processor forms a conservative estimate of expected advantages. It is important to note that non-blocking synchronization has been demonstrated to yield higher performance and efficiency on both legacy and single-core CPU packages as well. Therefore, potential benefits of the technology apply to both current and future system architectures and capacities.

4.2

Overview

The first two test articles represent the most used and the fastest growing web and application server products by web sites served, namely Apache and NGINX.

4.3

Performance and Efficiency Testing

In order to quantify web-server performance and energy consumption, metrics were developed as follows: For test purposes a single transaction represents one hypertext-transfer protocol (HTTP) request-response pair, as processed by a web-server. For each of the three web-servers a fixed number of requests were issued by a http client benchmarking program. Performance metrics are given in requests-per-second. The duration of time (counted in seconds) required to process a specified number of requests yields a transaction rate for each of the three web-servers tested. Energy efficiency metrics are obtained by measuring the energy consumed during each test. These measurements were accomplished through use of a "smart" power meter capable of recording, at one-minute intervals, the kilowatt-hours of energy consumed by a given device. The sum of these readings yielded an average rate of power and energy consumed by the server device during each phase of the test. Power efficiency is expressed in units of transactions-per-second-per-watt and energy usage in terms of joules-per-transaction.

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

8

4.3.1

Test Articles

The test articles studied in the test environment are:  Apache version 2.4.6;  NGINX version 1.6.2; and  Sequor version 0.9.9. 4.3.2

Test Environment

The test server hardware device was a Lenovo ThinkServer RS140, which utilizes a single-package, 2-core (4 hyper-threads) Intel CPU running at a clock speed of 3.4GHz. The server operating system was Centos version 7, a Red Hat Linux distribution. The test client benchmark applications utilized were two distinct, high-performance benchmarking programs: weighttp13 and httpress14. The former was employed in the HTTP benchmarking and the latter in the HTTPS (SSL) benchmarking. Note that although httpress does support both protocols, initial testing revealed that httpress can report occasional transaction failures. It is generally accepted that connection failures are to be expected when working with the httpress client due to issues with tuning of the underlying GnuTLS software and that this is not (necessarily) an issue with the web server being tested. Corresponding tests performed with weighttp recorded no transaction failures, thus it was decided to perform the HTTP tests with weighttp instead of httpress. 4.3.3

Test Method

Tests were conducted in such a manner as to ensure fairness to each web server. With the exceptions of the ‘screen’ remote shell assistant and the particular benchmark client program, no user processes were active during the tests. Regarding web server configuration, logging was disabled for all three servers due to the high transaction volume. NGINX was configured to use four worker processes instead of one, four being an optimal number for the server hardware device. For SSL benchmarking, each web server was configured as above with the additional setting that each web server employ a common cipher suite, namely RSA- AES128CBC-SHA (the mandatory cipher-suite of TLS version 1.1). The line-by-line details of the benchmark commands were encapsulated in a shell script, written to automate the procedure, as follows: 1. 2. 3. 4.

Synchronize the system clock with that of the power meter; Configure transient operating system settings; Start the particular web-server process; Wait at least one full minute and until the clock wraps to zero seconds;

13

http://redmine.lighttpd.net/projects/weighttp/wiki

14

https://bitbucket.org/yarosla/httpress/wiki/Home

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

9

5. Invoke the given benchmark client program with requisite parameters; and 6. Stop the web-server. Each of the above steps 3-6 were conducted for all three web servers. The server system time was captured immediately prior to and following step number 5. The entire process was repeated for both HTTP and HTTPS (SSL) transactions. Transient operating system settings were tuned in order to ensure high network stack throughput, specifically:     

Connection tracking on the loopback interface was disabled; The file/socket handle limit was increased to maximum; TCP/IP port reuse was enabled; TCP/IP connection recycling was enabled; and The restriction upon process locked physical memory pages was disabled.

Note that although the above connection availability settings are tuned to accommodate a high volume of non keep-alive connections, identical settings were applied to the keepalive tests for purposes of consistency. The client benchmark parameters were specified as follows:    

Utilize keep-alive connections (-k); Employ four threads, i.e. one per processor core (-t 4); Performance 256,000,000 transactions (-n 256000000); and Conduct 128 transactions concurrently (-c 128)

Note that HTTP transaction throughput is much greater than that of HTTPS so the number of HTTPS transactions was reduced from 256,000,000 to 64,000,000 in the interests of achieving output in similar time frames. Also note that the HTTPS transaction concurrency was reduced to 16. This was not done to alleviate server workload but rather to reduce the impact of httpress transaction failures15, which uses GnuTLS. Sequor Systems has developed its own SSL/TLS technology stack and did not experience SSL connection failures during testing. Given the power meter’s one-minute resolution per entry in the energy consumption log, each web server benchmark was set to commence on the minute, following a delay of no less than one minute. This synchronization simplified interpretation of the results and allowed for a more accurate extrapolation of the energy consumed during the final, fraction of a minute in which the individual web server benchmarks ran.

5 Results The Sequor web and application server produced the highest performance and efficiency numbers in both the HTTP and HTTPS tests, followed by NGINX. Apache came last in each test.

15

http://nmav.gnutls.org/

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

10

Note that because the Sequor web server processes a given transaction load in substantially less time than Apache or NGINX, Sequor uses dramatically less energy than Apache or NGINX. This is because Sequor takes better advantage of multi-core processing capacity and achieves dramatically higher transaction volumes per unit of energy used. During testing of the SSL/TLS protocols, both Apache and NGINX had connection failures with the httpress client. Note that during our testing, the Sequor web server did not experience any connection failures during HTTPS testing with httpress.

5.1

HTTP Results

Sequor demonstrated a performance factor of 8.39 over Apache and 3.53 over NGINX. This effectively means that it will take 9 servers running Apache or 4 servers running NGINX to achieve the same throughput as 1 server running Sequor. Apache demonstrated an efficiency of 13.36% compared to Sequor while NGINX demonstrated an efficiency of 29.67% compared to Sequor. The following tables summarize the results using the HTTP protocol:

HTTP

Total number of requests

Total time in seconds

Requests per second

Average power kilowatts

Total energy kilojoules

Joules per request

Apache

256,000,000

6,519.641

39,265.968

0.04811215

313.673923

0.00122529

NGINX

256,000,000

2,746.826

93,198.489

0.05142222

141.247879

0.00055175

Sequor

256,000,000

777.192

329,391.02

0.05392308

41.9085733

0.00016371

Power and time consumed to process 256 million HTTP requests Performance versus Apache

Performance versus NGINX

Performance versus Sequor

Efficiency versus Apache

Efficiency versus NGINX

Efficiency versus Sequor

Apache

1.000000

0.421316

0.119208

1.000000

0.450302

0.133606

NGINX

2.373518

1.000000

0.282942

2.220734

1.000000

0.296702

Sequor

8.388715

3.534296

1.000000

7.484720

3.370381

1.000000

HTTP

Comparison of HTTP performance and efficiency metrics

5.2

HTTPS Results

Sequor demonstrated a performance factor of 3.73 over Apache and 3.35 over NGINX. This effectively means that it will take 4 servers running Apache or NGINX to achieve the same throughput as 1 server running Sequor.

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

11

Apache demonstrated an efficiency of 27.44% compare to Sequor while NGINX demonstrated an efficiency of 29.76% compared to Sequor. The following tables summarize the results using the HTTPS protocol: HTTPS

Total number of requests

Total time in seconds

Requests per second

Average power kilowatts

Total energy kilojoules

Joules per request

Apache

64,000,000

2,430.060

26,336.798

0.05067500

123.143291

0.00192411

NGINX

64,000,000

2,185.942

29,277.995

0.05194444

113.547543

0.00177418

Sequor

64,000,000

652.044

98,152.885

0.05181818

33.7877345

0.00052793

Power and time consumed to process 64 million HTTPS requests

HTTPS

Performance versus Apache

Performance versus NGINX

Performance versus Sequor

Efficiency versus Apache

Efficiency versus NGINX

Efficiency versus Sequor

Apache

1.000000

0.899542

0.268324

1.000000

0.922077

0.274377

NGINX

1.111676

1.000000

0.298290

1.084509

1.000000

0.297565

Sequor

3.726834

3.352446

1.000000

3.644615

3.360614

1.000000

Comparison of HTTPS performance and efficiency metrics

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

12

5.3

Performance

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

13

5.4

Efficiency

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

14

6 Estimated Capital and Operating Cost Savings Based upon the findings above and numbers derived from market data and forecasts, we will attempt to calculate the minimum expected capital and operating costs savings for the average data center. The following electrical power commercial rate figures are taken from the Edison Electric Institute Typical Bills and Average Rates Report - Winter 2014 and the Canadian Electricity Association Key Canadian Electricity Statistics – June 2014. In both cases, the trend of the average rate was indicated as increasing. Key figures:  U.S. average commercial rate in cents per kilowatt-hour: 10.71  Canadian average commercial rate in cents per kilowatt-hour: 7.96 The following data center figures are taken from the key findings from Digital Realty’s 2013 Data Center Study of the North America Data Center Market based on research conducted by Campos Research & Analysis. Key figures:  Average data center IT load: 2.6 mega-watts  Average power density per rack: 8.5 kilo-watts The following assumptions are being made in order to calculate the typical number of active racks in the average data center: Assumptions:  Typical allocation of power to active racks: 66%  Typical number of active racks: 180 (2.6mW / 8.5 kW * 66%) The following estimates are being made in order to calculate the average number of nodes running web / application server software. Estimates:     

Estimated average number of U per rack: 42 Estimated average node allocation in rack: 2U Estimated average percentage of rack allocated: 66% Estimated average active 2U nodes per rack: 14 (42U / 2U * 66%) Estimated percentage of active nodes running web server software: 25%

Based upon the above, the typical number of active nodes running a web / application server application in the average data center is given by: Active Racks * Active Nodes per Rack * Active Nodes Running Web Server % The above yields: 180 * 14 * 25% = 630 Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

15

Based upon the above, the following tables estimate the capital and operating costs associated with provisioning each of the Apache, NGINX, and Sequor web / application server software packages.

6.1

Capital Cost Analysis

For purposes of estimating capital costs, a 2U server will be assumed to cost 1,000 dollars. It should be noted that many data centers may be required to double or triple this figure in order to approximate their actual costs. Additional typical data center capital costs will include infrastructure such as HVAC, racks, physical space, power supplies, etc. Software capital costs are as follows: Apache is free to license, NGINX is $1,350 per node per year, and Sequor is $1,000 per node per year. Costs restricted to server hardware and software licenses are as follows: Performance versus Apache

Performance versus NGINX

Performance versus Sequor

Number of servers required

Apache

1.000000

0.421316

0.119208

630

$630,000

$0

NGINX

2.373518

1.000000

0.282942

266

$266,000

$352,450

Sequor

8.388715

3.534296

1.000000

76

$76,000

$76,000

HTTP

Total hardware cost

Total software cost

Estimated Capital Cost Comparison of Apache, NGINX, and Sequor

HTTPS

Performance versus Apache

Performance versus NGINX

Performance versus Sequor

Number of servers required

Total hardware cost

Total software cost

Apache

1.000000

0.899542

0.268324

630

$630,000

$0

NGINX

1.111676

1.000000

0.298290

567

$567,000

$751,275

Sequor

3.726834

3.352446

1.000000

170

$170,000

$170,000

Estimated e-Commerce Capital Cost Comparison of Apache, NGINX, and Sequor

Based upon above assumptions, conservative estimates of capital costs running Apache, NGINX, and Sequor web / application server software are as follows:  Apache:  NGINX:  Sequor:

$630,000 $618,450 $152,000

Based upon above assumptions, conservative estimates of capital costs running Apache, NGINX, and Sequor as an e-commerce solution are as follows:  Apache:

$630,000

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

16

 NGINX:  Sequor:

6.2

$1,318,275 $340,000

Operating Costs Analysis

For purposes of estimating operating costs, price in cents per kilo-watt hour is assumed to be 10 cents. For simplicity purposes, energy consumption for a 2U server is assumed to be 1,000 kilo-watt hours per year. It should be noted that many data centers may be required to double or triple these figures in order to approximate their actual costs. Additional typical data center operating costs will include management, maintenance, and support costs. Support related costs are as follows: Apache support costs vary, NGINX is a minimum of $1,350 per node per year, and Sequor is $1,250 per node per year. Support for both NGINX and Sequor is optional and actual costs may vary. Costs restricted to electrical power use and software support are as follows: Efficiency versus Apache

Efficiency versus NGINX

Efficiency versus Sequor

Apache

1.000000

0.450302

0.133606

630

$63,000

NGINX

2.220734

1.000000

0.296702

284

$28,400

Sequor

7.484720

3.370381

1.000000

85

$8,500

HTTP

Number of servers required

Total electrical costs

Total support costs Varies by support level Varies, up to $376,300 Varies Up to $106,250

Estimated Operating Cost Comparison of Apache, NGINX, and Sequor

HTTPS

Efficiency versus Apache

Efficiency versus NGINX

Efficiency versus Sequor

Number of servers required

Total electrical costs

Apache

1.000000

0.922077

0.274377

630

$63,000

NGINX

1.084509

1.000000

0.297565

581

$58,100

Sequor

3.644615

3.360614

1.000000

173

$17,300

Total support costs Varies by support level Varies, up to $769,825 Varies, up to $216,250

Estimated eCommerce Operating Cost comparison of Apache, NGINX, and Sequor Based upon above assumptions, conservative estimates of operating costs running Apache, NGINX, and Sequor web / application server software are as follows:  Apache:  NGINX:  Sequor:

Lowest cost: $63,000 Lowest cost: $28,400 Lowest cost: $8,500

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

17

Based upon above assumptions, conservative estimates of operating costs running Apache, NGINX, and Sequor as an e-commerce solution are as follows:  Apache:  NGINX:  Sequor:

6.3

Lowest cost: $63,000 Lowest cost: $58,100 Lowest cost: $17,300

Total Cost of Ownership

Based upon the above, server rationalization by performing software upgrades to the Sequor web and application server is expected to achieve expected benefit factors of:  4x improvement in capital costs over Apache for delivery of HTTP content  4x improvement in capital costs over NGINX for delivery of HTTP content  2x improvement in capital costs over Apache for delivery of HTTPS content  4x improvement in capital costs over NGINX for delivery of HTTPS content  7x improvement in operating costs over Apache for delivery of HTTP content  3x improvement in operating costs over NGINX for delivery of HTTP content  4x improvement in operating costs over Apache for delivery of HTTPS content  4x improvement in operating costs over NGINX for delivery of HTTPS content

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

18

7 Conclusions A new technique for overcoming performance and efficiency issues is the use of “nonblocking synchronization”. The Sequor web and application server utilizes this novel technique and Sequor’s patented non-blocking algorithms to achieve dramatically higher levels of performance and efficiency than those offered by either Apache or NGINX. The performance of the Sequor web server conveys important benefits in the data center, including improved efficiency, scalability, availability and capacity. These benefits foster the potential for dramatic savings on capital and operating costs, as well as improved avenues for business capacity and growth. Additionally, Sequor’s secure (SSL/TLS) web and application server platform provides industry leading protection against data theft while mitigating issues related to bandwidth starvation and network and denial of service attacks. The Sequor web and application server software represents a compelling solution to performance and efficiency issues facing cloud, enterprise, and data center businesses.

8 Appendices 8.1

Test Configuration Hardware Specifications

8.1.1

Meter

Manufacturer

Model Number

Firmware Version

Specifications

ENLOGIC

EZ1254

9.0e

Inline energy meter, single phase, 50/60 Hz

Manufacturer

Model Name

Model Number

Specifications

Lenovo

ThinkServer

RS140

1U Intel Core i3-4130 3.40 GHz, 4GB DDR3 RAM

8.1.2

Server

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

19

8.2

Test Configuration Software Specifications

8.2.1

Operating System

Manufacturer

Operating System

Software Version

Specifications

Red Hat

CentOS

7.0.1406

Linux version 3.10.0 gcc version 4.8.2

Manufacturer

Web Server

Software Version

Specifications

Apache

Apache

2.4.6

NGINX

NGINX

1.6.2

Sequor Systems

Sequor

0.9.9

8.2.2

8.2.3

Web Server

Web Clients

Manufacturer

Web Client

Software Version

Specifications

Lighty labs

Weighttp

0.3

build-date: Sep 27 2012 14:10:15

Yaroslav Stavnichiy

Httpress

1.1

build-date: Jan 21 2015 23:56:48

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

20

8.3

About the Authors

Sequor Systems Sequor Systems is an independent software vendor specializing in high performance and high efficiency software systems. These systems enable the Information Technology sector to realize the performance and efficiency benefits of modern technology in today’s mobile, desktop, server, cloud, and high performance computing environments. The authors wish to extend their gratitude to Internetworking Atlantic, Inc. and Rogers Telecommunications Inc for provisioning of equipment and hosting services used in producing this white paper.

8.4

Further Reading

Performance Scalability of a Multi-core Web Server, 2007 Bryan Veal, Annie Foong Calculation of TCO for Energy, IBM Systems Magazine, November 2011, Scott Barielle The Art of Multiprocessor Programming, 2008 Maurice Herlihy, Nir Shavit Why Intel is designing Multi-Core Processors, 2006 Geoff Lowney The Cloud Wars Part V: The New Cloud Nine, May 2013, Bank of America, Merrill Lynch

Data Center Performance and Efficiency Challenges: Disruptive Software Solutions

21