Accelerating Business Intelligence with Large-Scale System Memory A Proof of Concept by Intel, Samsung, and SAP Executive Summary Real-time business intelligence (BI) plays a vital role in driving competitiveness across industries. Transforming operational data into actionable information that drives smart business decisions has become a core requirement. As a result, choosing the right solution building blocks to optimize BI infrastructure for performance and power consumption has become a strategic imperative in its own right. To identify the value of very high amounts of system memory to BI implementations, this proof of concept study tests SAP InMemory Appliance Software (SAP HANA ) on servers based on the Intel Xeon processor E7 family and provisioned with next-generation memory from Samsung, a pioneer in producing 32-GB DIMMs and DIMMs based on 30nm-class process technology. Solution provided by:
This paper begins with a review of the role of high-performance, realtime BI to the enterprise, before describing the test procedure and workloads used to compare the various memory configurations under test. It then presents the test results and provides an analysis of those results that shows their value to real-world implementations. The paper is intended primarily for IT decision makers to help them understand the roles of various memory configurations in BI systems. It will also be useful to end users, and sales and marketing teams involved with BI infrastructure. Core Findings Full-text-search workload benefits from both 32-GB DIMMs and 30nm-class process technology: Throughput per GB of loaded data increased by more than 1.4x as data sizes were increased. 30nm-class DIMMs provided up to 20 percent power reduction on the system level, relative to 50nm-class DIMMs. Analytical (SAP-H) workload cannot benefit from 32-GB DIMMs but benefits from 30nm-class process technology: Similar runtimes were achieved with all memory configurations tested. 30nm-class DIMMs provided approximately 20 percent lower average power consumption than 50nm-class DIMMs. Changing Massive Stores from an Asset to a Liability Organizations of all types and sizes maintain increasingly large data stores, which are widely recognized as a significant potential asset to business strategy and operations. Unfortunately, the ability to turn that potential into an actual benefit is typically limited by the ability of conventional data-analysis and information retrieval systems to handle large-scale business data with the speed and flexibility required. Data analysis typically depends on custom reports that IT develops and runs in batches, often during offhours. Overcoming that limitation enables business users to respond to trends as they emerge, improving the value of the data by enabling betterinformed business decisions. Real-time data analysis enables benefits that are broadly applicable across usage areas, including the following: Order management and supply chain logistics. Efficient supply and demand planning, volume management, transportation, and pricing Manufacturing and production. Improved planning, volume management, asset and personnel utilization, and performance management Operational intelligence and efficiency. Optimized sales and marketing analytics, trade promotion management, budgeting, and IT strategy Realizing those benefits requires a highquality, integrated solution stack such as the one used in the proof of concept tested in this paper. It is based on the three building blocks described below. Server Platform: Intel Xeon E7 Processor Family The Intel Xeon processor E7 family delivers scalable performance with up to 10 processor cores per socket, enabled for two software threads per core by Intel Hyper-Threading Technology, to deliver high parallelism for transactional workloads. A 30- MB last-level cache keeps a large data store close to the processor, for access with minimal latency. An advanced, integrated memory controller supports 32-GB DIMMs that enable as much as 2 TB of RAM to be provisioned on a four-socket system. In addition, Intel Intelligent Power Technology adjusts power usage dynamically to the workload, for lower operating expense, and advanced reliability features enhance the processor s suitability for mission-critical enterprise BI implementations. 2 Accelerating Business Intelligence with Large-Scale System Memory
In-Memory Data-Analysis Software: SAP HANA SAP HANA, which is highly optimized for the Intel Xeon processor E7 family, delivers order-of-magnitude performance improvements that enable business users to sift through massive amounts of information, create rich, ad hoc reports, and analyze operational data in real time. The latency associated with reading data from disk before scanning or search operations and calculations, and writing it back afterward, tends to limit the performance of conventional database systems. SAP HANA overcomes that limitation by using main memory as the primary data store, building on the tradition of in-memory computing begun with SAP NetWeaver Text Retrieval and Classification (TREX) and Business Warehouse Accelerator (BWA). Current BI applications take advantage of the structured parts of business data, leaving untapped an important source of information: comments, tags, long text fields, and documents. Future decision-support systems will stand out by providing analytical as well as full-text-search capabilities, in a blended, seamlessly integrated fashion. One of SAP HANA s core components, the in-memory columnar database, exposes highly optimized data structures to its calculation and search engine. This paper demonstrates that full-text-search queries on SAP HANA benefit dramatically from large memory footprints, such as those enabled by Samsung s 32-GB DIMMs. System Memory: Samsung 32-GB and 16-GB 30nm-class DIMMs As an industry leader in memory manufacturing, Samsung is the only provider of mass-produced DIMMs available in sizes up to 32 GB per module. That factor is necessary to realize the full memory capacity of 2 TB in a four-way server based on the Intel Xeon processor E7 family. The company is also unique in using 30nm-class process technology for large-scale manufacturing of memory modules. The extremely small feature sizes associated with that process technology enable high performance with low power consumption. Building a Next-Generation Test Environment In-memory computing depends on the ability to hold very large data sets in RAM, which in turn suggests that systems benefit from being provisioned with very large amounts of system memory. A series of factors have enabled that requirement to be met in mainstream platforms, and they include the following: Advanced server memory controllers and 64-bit platforms enable systems to utilize very high amounts of RAM. Table 1. Components of the test system. Processors Intel Xeon Processors E7-4870 @ 2.4GHz Memory Modules Samsung 30nm-class 4 Gb 32 GB (Part #: M393B4G70BM0-YH9) Samsung 30nm-class 2 Gb 16 GB (Part #: M393B2K70DM0-YH9) Samsung 50nm-class 2 Gb 16 GB (Part #: M393B2K70BM1-CF8) High-capacity memory modules allow for large memory footprints to be achieved with a reasonable number of DIMMs. Memory with low power requirements makes large amounts of memory viable, without excessive energy use and waste-heat generation. To explore the behavior of mainstream systems configured with very large amounts of system memory, test engineers from SAP, Intel, and Samsung collaborated on this study. Note: The test configuration used in this study does not correspond to one that would be usable in a production environment. Rather, this research examines theoretical behavior using very large amounts of system memory. The current version of SAP HANA is certified for use with 512 GB of memory, as opposed to the 2 TB used as the upper limit in this study. The system under test was a four-way Intel white box server, configured with the components shown in Table 1, running SUSE* Linux* Enterprise Server 11. Business Intelligence Software SAP In-Memory Appliance Software (SAP HANA ) Development Version Accelerating Business Intelligence with Large-Scale System Memory 3
Trials tested the impacts of memory capacity (1 TB per server versus 2 TB per server), module size (16 GB versus 32 GB per DIMM), and process technology (30nm class versus 50nm class) on performance and power consumption. By manipulating server BIOS settings, testing was also conducted to compare performanceoptimized settings with poweroptimized ones. The workloads employed included the following: Full-text-search workload. Each element of this set of phrase queries consisted of several words that, to be considered a hit, had to occur as a sequence within the test data. The workload consisted of several query logs that simulated real-world user input; this testing used four parallel user processes that each initiated queries in sequence. Data sets (loaded text-parts) used in this testing ranged in size from 220 GB to 1.32 TB. Because of the large data sets used with this workload and the relatively low level of calculation performed on that data, these operations tended not to be processor bound. Therefore, the fulltext-search workload benefits from the very high maximum memory capacities made possible by using Samsung s 32-GB DIMMs and the Intel Xeon processor E7 family. Analytical workload (SAP-H). This set of queries is similar to the TPC-H transactional decisionsupport benchmark, adapted to a typical SAP enterprise resource planning schema. The data set used in this testing, designated as SF 100 consisted of loaded data that consumed approximately 550 GB of main memory, and the query set consisted of 20 different queries running several times, simulating one, 15, and 20 concurrent users. Because of the large degree of calculation performed on the data by this workload, these operations tended to be processor bound. Driving Successful Real-Time BI with Memory Innovation Test results revealed insights through both the presence and absence of significant differences in performance and power consumption. The fulltext-search workload benefited from the higher level of system memory (2 TB versus 1 TB). Both workloads achieved lower power usage from the memory based on 30nm-class process technology, relative to the memory based on 50nm-class process technology. SAP HANA Full-Text-Search Workload Throughput per GB of Loaded Data (2-TB, 30nm-Class Memory Configuration) Throughput per GB of Loaded Data 1.00x 220 1.23x Full-Text-Search Workload Results As noted above, the largest data set used in testing with the full-text-search workload was 1.32 TB. Therefore, the 2-TB memory configuration (but not the 1-TB configuration) enabled the system to keep all of the data in memory at the same time. Testing with the full-text-search workload revealed that, as the overall size of the loaded data set increases, throughput per GB of loaded data increases, as shown in Figure 1. This result indicates that the full-text-search functionality is able to benefit from the high ratio of memory to processing resources in the system under test. Note that the values in all charts in this paper are expressed as relative to a baseline value within the set of results that is defined for that purpose as 1.0x. 440 880 1320 Size of Loaded Data (GB) Figure 1. The full-text-search workload generates higher throughput per GB of loaded data as the overall size of the loaded data set increases. 1.36x 1.42x 4 Accelerating Business Intelligence with Large-Scale System Memory
Average power consumption of the full-text-search workload on a system configured with 1 TB of system memory with 16-GB DIMMs is lower by approximately 20 percent when using modules manufactured with 30nm-class process technology versus those manufactured using 50nm-class process technology, as shown in Figure 2. This result is found consistently with various sizes of loaded data. Moreover, the average power consumption with a 2-TB memory configuration is only about four percent higher than it is with a 1-TB configuration. Average Power Consumption 1.00x 0.96x 220 2 TB (30nm Class) SAP HANA TM Full-Text-Search Workload Average Power Consumption 1.17x 1.03x 1.01x 440 1.24x Size of Loaded Data (GB) 1 TB (30nm Class) 1.08x 1.05x 880 1.27x 1 TB (50nm Class) To gauge the effect of BIOS settings made to optimize the system for power usage or performance, performance per watt was compared using a 2-TB, 30nm-class memory configuration under various conditions, as summarized in Figure 3. Regardless of the size of loaded data being tested, the performance-optimized case demonstrated higher performance per watt than the power-optimized case, even though the average power consumption was lower in the poweroptimized case. Figure 2. At various sizes of loaded data, the full-text-search workload uses approximately 20 percent lower average power with 30nm-class memory than with 50nm-class memory. Throughput per Wattt 1.00x SAP HANA TM Full-Text-Search Workload Performance per Watt (2-TB, 30nm-Class Memory Configuration) 220 0.96x 1.19x 1.30x 1.27x 440 880 1320 Size of Loaded Data (GB) 1.37x 1.30x 1.39x Power Optimized Performance Optimized Figure 3. Performance-optimized BIOS settings provide approximately a 7 to 10 percent increase in performance per watt versus power-optimized settings. Accelerating Business Intelligence with Large-Scale System Memory 5
Analytic (SAP-H) Workload Results Testing of power usage by the SAP-H analytic workload was gauged by measuring the average power consumption while running the workload under different memory configurations and various numbers of simulated concurrent users, as shown in Figure 4. The results show average power consumption in the 1-TB configuration based on 16-GB 30nmclass DIMMs to be approximately 20 percent lower than that of the 1-TB configuration based on 16-GB 50nmclass DIMMs. Average Power Consumption 1.00x 0.97x 1 SAP HANA TM Analytic (SAP-H) Workload Average Power Consumption 1.21x 2 TB (30nm Class) 1.44x 1.42x 15 1.78x Number of Simulated Concurrent Users 1 TB (30nm Class) 1.45x 1.42x 20 1.84x 1 TB (50nm Class) Figure 4. Power consumption by the SAP-H workload was approximately 20 percent lower using 30nm-class memory than it was using 50nm-class memory when the system was configured with 1 TB of RAM. 6 Accelerating Business Intelligence with Large-Scale System Memory
Analysis and Conclusions The most significant result of this testing is the ability of the full-text-search workload to benefit from the large memory footprint. The raw volumes of loaded data used up to 1.32 TB for the full-text-search workload follows real-world circumstances, where full-text search is commonly applied to very large data sets, so efficiency requires the use of very large indexes. Because such indexes represent pre-calculated intermediate results, larger amounts of available system memory mean more intermediate results can be stored for when they are needed. Greater availability of such indexes corresponds to greater presence of pre-calculated results, which drives faster response times. Thus, the use of indexes, as with the full-text-search workload, directly benefits from the presence of large memory footprints, since more memory allows for larger indexes. The system used in this testing represents the first known system configured with 2 TB of memory to be tested with SAP HANA, a configuration that is uniquely made possible using 32-GB DIMMs. The other most significant test result was the significantly lower power consumption by memory modules based on 30nm-class process technology, compared with those based on the older 50nm-class process technology, which is most common within the industry. Because lower power consumption equates to lower operating expense, the value of the smaller process technology is clear. As an industry leader and the only mass provider of 32-GB DIMMs and DIMMs based on 30nm-class process technology, Samsung provides customers with mature products for large memory footprints. That capability extends the broader benefits of using SAP HANA on the Intel Xeon processor E7 family. The combination of industry leadership from Intel, SAP, and Samsung delivers cutting-edge innovation that equates to enhanced capabilities with lower operating costs. Accelerating Business Intelligence with Large-Scale System Memory 7
Learn more about the Intel Xeon processor E7 family at www.intel.com/xeon Learn more about SAP HANA at www.sap.com/hana Learn more about Samsung memory products at www.samsung.com/greenmemory SAMSUNG ELECTRONICS RESERVES THE RIGHT TO CHANGE PRODUCTS, INFORMATION AND SPECIFICATIONS WITHOUT NOTICE. Products and specifications discussed herein are for reference purposes only. All information discussed herein is provided on an AS IS basis, without warranties of any kind. This document and all information discussed herein remain the sole and exclusive property of Samsung Electronics. No license of any patent, copyright, mask work, trademark or any other intellectual property right is granted by one party to the other party under this document, by implication, estoppel or other-wise. Samsung products are not intended for use in life support, critical care, medical, safety equipment, or similar applications where product failure could result in loss of life or personal or physical harm, or any military or defense application, or any governmental procurement to which special terms or provisions may apply. For updates or additional information about Samsung products, contact your nearest Samsung office. All brand names, trademarks and registered trademarks belong to their respective owners. 2011 Samsung Electronics Co., Ltd. All rights reserved. INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL PRODUCTS. NO LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY INTELLECTUAL PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL S TERMS AND CONDITIONS OF SALE FOR SUCH PRODUCTS, INTEL ASSUMES NO LIABILITY WHATSOEVER, AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO SALE AND/OR USE OF INTEL PRODUCTS INCLUDING LIABILITY OR WARRANTIES RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT, COPYRIGHT OR OTHER INTELLECTUAL PROPERTY RIGHT. UNLESS OTHERWISE AGREED IN WRITING BY INTEL, THE INTEL PRODUCTS ARE NOT DESIGNED NOR INTENDED FOR ANY APPLICATION IN WHICH THE FAILURE OF THE INTEL PRODUCT COULD CREATE A SITUATION WHERE PERSONAL INJURY OR DEATH MAY OCCUR. Intel may make changes to specifications and product descriptions at any time, without notice. Designers must not rely on the absence or characteristics of any features or instructions marked reserved or undefined. Intel reserves these for future definition and shall have no responsibility whatsoever for conflicts or incompatibilities arising from future changes to them. The information here is subject to change without notice. Do not finalize a design with this information. The products described in this document may contain design defects or errors known as errata which may cause the product to deviate from published specifications. Current characterized errata are available on request. Contact your local Intel sales office or your distributor to obtain the latest specifications and before placing your product order. Copies of documents which have an order number and are referenced in this document, or other Intel literature, may be obtained by calling 1-800-548-4725, or by visiting Intel s Web Site http://www.intel.com/. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark* and MobileMark*, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more information go to http://www.intel.com/performance. *Other names and brands may be claimed as the property of others. Copyright 2011 Intel Corporation. All rights reserved. Intel the Intel logo, and Xeon are trademarks of Intel Corporation in the U.S. and other countries. 0911/SM/MESH/PDF 326058-001US