White Paper. A Validation and Comparison of Storage Efficiency Technology: Compression. IBM Real-time Compression Appliance EMC VNX Compression

Transcription

1 Fifth Fifth Avenue, 7th 7th Floor New York, Floor NY New York, NY White Paper A Validation and Comparison of Storage Efficiency Technology: Compression IBM Real-time Compression Appliance EMC VNX Compression

2 Printed in the United States of America. Copyright 2012 Edison Group, Inc. New York. Edison Group offers no warranty either expressed or implied on the information contained herein and shall be held harmless for errors resulting from its use. All products are trademarks of their respective owners. First Publication: January 2012 Produced by: Manish Bhardwaj, Senior Analyst; Renee Boucher Ferguson, Author; Manny Frishberg, Editor; Barry Cohen, Editor-in-Chief

3 Table of Contents Executive Summary... 1 Introduction... 2 Objective... 2 Audience... 2 Overview... 3 Technology Solutions... 5 EMC Technology... 5 IBM Real-time Compression... 6 Why the IBM Real-time Compression Appliance is Unique... 9 Set It and Forget It... 9 Transparency... 9 Performance Compression Research Results Test Environment SUT Installation Test Objectives Test One File Copy File Copy Time and Compression Levels System Utilization Results What the Test Results Mean Test Two Compression over Time What the results mean Test Three IOzone with Large Files What the Test Results Mean Test Four Database Performance Measured with TPC-C What the Test Results Mean Test Five Backup using NDMP What the Test Results Mean Conclusion... 25

4 Appendices Test Setup Servers: Software: Storage: Network: A Note on the Use of Benchmarks Test Results Test Case 1 File Copy Test Case 2 Compression Over Time Test in Detail Test Case 3 IOzone with Large Files Test in Detail Test Case 4 TPC-C Test in Detail Test Case 5 Backup using NDMP Test in Detail... 34

5 Executive Summary In 2010 IBM launched its Real-time Compression Appliance (RtCA), bringing the first real-time data compression technology for Network Attached Storage (NAS) to market. By compressing data before it is stored, real-time compression enables a tremendous increase in existing storage capacity and gives end users an opportunity to optimize their existing storage footprints. This white paper validates the claims made by IBM regarding its RtCA and demonstrates the technology as compared to EMC s compression technology embedded in the VNX system. To illuminate IBM s technical lead with its data compression technology, Edison Group compared the differences between the IBM RtCA and EMC VNX Series of storage systems, a midrange storage array with embedded compression. Edison set up five separate tests, from high performance database to file copy, in order to validate four key concepts of RtCA: performance, maximum compression ratios, overall storage savings, transparency and ease of implementation. In the course of our testing, Edison found outstanding results in these areas, including: Edison found no negative impacts on performance with RtCA inline Edison found a performance improvement in areas such as backup and IOzone, where system throughput was almost 57 percent greater for read operations with the RtCA compression enabled than without compression. System database performance tested via TPC-C workloads were not affected by the use of the RtCA, which delivered 64 percent data compression The RtCA showed a maximum file-system compression ratio of 91 percent throughout the life of the file, compared to a VNX compression of only 4.46 percent. In standard file copy, real-time ccompression delivered 91 percent compression while lowering network utilization by 200 percent and filer CPU overhead by over 30 percent. Backup via the RtCA, over the NDMP protocol using Symantec NetBackup, required 53 percent less time than backing up without compression. It required 55 percent less time than backing up with VNX deduplication/compression. Implementation for the RtCA was essentially a set it and forget it experience. We found, in particular that RtCA is completely transparent to IT processes, requiring no additional IT administrator resources to run compression on primary data. Edison: Array Based Compression Storage Efficiency Technology Page 1

6 Overall Edison found that average system performance of the storage array with Realtime Compression enabled was equal to or better than performance without Real-time Compression. It was almost always better for actively used, primary production data than array-based compression on the EMC VNX In summary, the test results prove that IBM RtCA provided a seamless implementation experience, resulting in superior compression in primary data environments, as compared to its competitor, the EMC VNX Series. Introduction Objective IBM has asked Edison to perform product testing on its Real-time Compression technology and EMC VNX embedded compression, to validate its claims in the areas of maximum compression, performance and transparency to applications, storage, networks, and downstream processes. Audience The competitive white paper is a public report that offers an objective third-party perspective, provides evidence that validates claims, and educates IT decision makers, storage and other data administrators on strengths and benefits of IBM s Real-time Compression. Edison: Array Based Compression Storage Efficiency Technology Page 2

7 Overview The fact that data is growing exponentially is no secret. IDC expects the volume of digital data to grow to 2.7 zetabytes in 2012 that s 2.7 trillion gigabytes of data up 48 percent from More than 90 percent of the information generated will be unstructured data from all manner of sources, including articles, s, social media, audio, image and video data. To compensate for this data explosion, over the next decade the number of servers (virtual and physical) worldwide will grow by a factor of 10. The amount of information managed by enterprise data centers will grow by a factor of 50 and the number of files data centers will have to deal with will grow by a factor of 75. Meanwhile, the number of IT professionals in the world will grow by just 150 percent in the same time frame. 1 To say that the rapid propagation of data presents a backup and storage issue for the world s IT organizations is an understatement. While the cost of physical data storage has come down significantly according to some sources, cost per GB is dropping at 25 percent per year 2 those cost savings are offset by capacity growth. The constant influx of data still requires a massive build-out of infrastructure both physical capacity and technology to handle not only the growth of capacity but also to manage critical backup and recovery requirements. This at a time when IT budgets are stagnant or only increasing incrementally. To put this in context, data capacity is growing 40 to 60 percent annually (a result of both the unstructured data explosion and regulatory requirements to keep data longer) 3 yet IT spending is expected to increase at only 5 percent annually. 4 IT organizations typically approach the problem of ever-increasing data storage needs by implementing technology to shrink the amount of data that is actually stored. There are two predominant methods to data shrinkage today: deduplication and compression. Both do what their names imply, reducing the amount of data that is actually stored. With compression, the same amount of data is transmitted by using a smaller number of bits, typically utilizing the industry standard Lempel-Ziv compression algorithm for lossless compression. Deduplication takes out duplicate copies of the same data. For 1 The 2011 Digital Universe Study: Extracting Value From Chaos, IDC A good quote on this topic, from Wayne Salpietro at Database Trends and Applications can be found at: 3 Data Growth Remains IT s Biggest Challenge Gartner Says, Computer World, Lucas Mearian, November 2, _s_biggest_challenge_gartner_says 4 Big data: The Next Frontier for Innovation, Competition, and Productivity, McKinsey Global Institute, June 2011 Edison: Array Based Compression Storage Efficiency Technology Page 3

8 example, replacing a string of 10 repeated digits with a command to repeat the digit 10 times. Deduplication is optimally used for inactive data such as backup files and archives where data tends to be repeated, but it is not a good fit for active, primary data sets. To be effective and help mitigate the impacts of expanding data requirements, new storage technologies need to lessen the impact exponential data influx. At the same time, they should be transparent to operations, with no impact to any of the characteristics behind IT purchases of storage in the first place. IBM, with its Real-time Compression technology, has developed a method for shrinking primary data in real time, before it hits the storage array. The Real-time Compression Appliance (RtCA) is transparent to infrastructure and process, and there is no impact to performance. In fact, hands-on testing has proved an actual increase in performance in some cases. Given that there is no impact on performance, data compression can be used on primary data for the first time. IBM s Real-time Compression Appliance sits at the front of a company s NAS (Network Attached Storage) and processes primary data coming into and out of the NAS. Because data is compressed on initial write, primary storage is optimized. The result is not only a tremendous up-front savings, but also a tremendous cascade effect throughout the data lifecycle. Because IBM s Real-time Compression technology has the ability to cut the amount of data that is stored by 50 percent or more compressing data in real time before it hits the storage array IT organizations are able to realize the expected savings of lower storage costs. The bottom line is that, because there is less capacity used on primary storage, there is a cascade effect through the data lifecycle. This results up-front storage savings lower capacity requirements mean less physical storage. In turn, this creates downstream savings on process and secondary storage by allowing you to put more snapshots in the same space to create better Recovery Time Objectives (RTOs) and Recovery Point Objectives (RPOs) 5, and lowering your replication and backup resource requirements. This, of course, has a positive impact on long-term storage savings and utilities costs. Ultimately, by using IBM Real-time Compression technology, companies are able to mitigate the impacts of data proliferation, increasing storage requirements and rising storage expenditures regardless of changes in hardware price. 5 RPO (Recovery Point Objective) and RTO (Recovery Time Objective) are parameters associated with data restoration. RTO refers to the time an organization can operate without a specific application. RPO refers to the maximum allowable data loss how much data can an organization afford to lose, Edison: Array Based Compression Storage Efficiency Technology Page 4

9 Technology Solutions EMC Technology EMC s VNX Series unified storage system, launched in January 2011, is designed to replace the company s aging CLARiiON and Celerra midrange storage systems. The two product lines were essentially brought together to create a unified storage platform optimized for virtualized IT (EMC is, after all, the parent company to VMware). CLARiiON brings a SAN-based block storage platform to VNX that supports Fibre Channel and iscsi. Celerra on the other hand brings CIFS and NFS capabilities with filelevel deduplication and compression to the VNX Series. The VNX Series includes both the VNX and VNXe platforms, and includes five separate product lines, ranging from VNX 5100 to VNX VNX 5300 and above are designed for midrange to enterprise-class deployments, while VNXe the 5100 series is designed primarily for small (and mid-sized) business deployments. For this comparison of compression capabilities Edison utilized the EMC VNX 5500, which is geared toward larger midrange- and enterprise-class deployments. EMC documentation uses the terms compression and deduplication when describing capacity optimization on the VNX. In the current context, this can be confusing because the compression that results from deduplication is different from what is occurs on other systems offering compression, and from that offered by the IBM RtCA. VNX performs all deduplication and compression processing as a background, asynchronous operation that acts on file data after it is written into the file system. Because of this fact, EMC VNX deduplication and compression is geared toward stale data and not active data. This is a major differentiator from IBM RtCA, which targets primary, active data for compression. According to EMC s instructional video 6, VNX data compression is intended for relatively inactive data files, including: Backup copies Archives Static data repositories Other redundant data 5 Block Data Compression video Edison: Array Based Compression Storage Efficiency Technology Page 5

10 The key rational behind utilizing data compression on non-active data is that there is a performance impact when using compression. Therefore, EMC recommends that compression be utilized on relatively inactive data, which has lower performance requirements. IBM Real-time Compression IBM Real-time Compression is based on the industry standard LZ compression algorithm for lossless compression. It utilizes a table-based compression model in which table entries (typically called a dictionary) are substituted for repeated strings of data, improving both the compression ratio and decompression speed. Where the innovation with Real-time Compression comes in is IBM s 35 patents that make lossless compression in real time possible. The patented IBM technologies that define how to make LZ compression operate in real time reside in the RACE engine. RACE takes incoming data streams and compresses the data on write, leaving the metadata of the file intact. All write commitments come directly from the array so that the RtCA is not affecting the high availability built into the infrastructure. Additionally, all of the metadata of the file is preserved in the compressed file so the application believes it is reading the original file. 7 IBM Real-time Compression differs from traditional LZ compression algorithms in several ways. First, with traditional compression, the original data is treated as fixed input which, when compressed, will vary in layout, as shown in Figure 2 below. 8 Figure 1: Traditional Compression Technology 7 IBM Random Access Compression Engine RACE, IBM white paper, December 7, Graphics depicting RtC technology features: courtesy of IBM. Edison: Array Based Compression Storage Efficiency Technology Page 6

11 If the contents of the compressed file are changed, the file can become fragmented, as shown in Figure 3. When this occurs, the file will become larger to accommodate the fragments. Figure 2: Traditional Compression over Time IBM RtC works differently. As can be seen in Figure 4 below, the RtC technology treats data as an I/O stream, a variable input. That stream is compressed into fixed length blocks of 32K for NFS, a configurable 61.44K for and CIFS (Windows). A map keeps track of what data is where. Figure 3: Real-time Compression Technology When data is changed, this fixed block size and data map prevents fragmentation and thus file size growth. As can be seen in Figure 5, the map is updated with the change, the data block that has been changed is flagged and an update block contains the changed data. When another update occurs, the flagged block is replaced with the update block, the map is once again updated and the file size remains unchanged. Edison: Array Based Compression Storage Efficiency Technology Page 7

12 Figure 4: Real-time Compression over Time Another difference between IBM RtC and traditional compression is that traditional compression is location-based: compression is based upon where in the file the data resides. RtC utilizes time-based compression also, tracking data over time. As illustrated in Figure 6, location-based compression cannot find like objects at different locations in the file, whereas with time based compression, all of the like objects can be compressed together, yielding a 10 percent to 20 percent more efficient compression ratio. Figure 5: Location-based versus time-based compression Edison: Array Based Compression Storage Efficiency Technology Page 8

13 Why the IBM Real-time Compression Appliance is Unique Set It and Forget It The Real-time Compression Appliance installs in several quick steps that can be completed in a few hours 9. Because Real-time Compression is essentially a plug-andplay appliance, two primary characteristics emerge from the run-time experience: 1. IT managers are able to implement Real-time Compression quickly, with minimal configuration effort and no end user down time. This means that though adding an RtCA to the storage network does add an additional device to manage, that addition is relatively simple and requires minimal post-installation configuration Because the appliance is invisible to the server, there is little additional administrative overhead to use the RtCA once the system is configured. 10 Transparency Use of the RtCA is completely transparent from the perspective of the storage arrays and hosts making use of the technology. Once installed and enabled on the storage network hosts, applications and arrays communicate with no configuration changes. Users viewing files on a server are even unaware that a file has been compressed; the operating system displays the uncompressed file size. This prevents confusion when copying files that might occur if the original and destination file sizes were reported as being different. In addition, all of the existing downstream processes stay the same. Snapshots work as they did before. More snapshots can be put in the same amount of space as the old snapshots took, allowing for improved RPO/RTO. Replication works as it did before only now it utilizes bandwidth and remote storage more efficiently. Backups work as they did before, but they operate faster and require less capacity. 9 Take a look at this YouTube presentation for a hands-on demo of the Real-time Compression Out of the Box Experience: ( 10 There is some additional management overhead required for use of the RtCA when compared to using deduplication on the VNX. On VNX, an administrator can enable deduplication by clicking a checkbox when a file system export is created; a very simple process. When the RtCA is used, the administrator must identify exported file systems and click the option for applying compression. Edison has not performed a formal complexity comparison of these approaches, but we believe that the RtCA balances the apparent greater management complexity by providing greater flexibility by enabling compression across files, shares and volumes vs. at a volume level, which is the VNX approach. Edison: Array Based Compression Storage Efficiency Technology Page 9

14 Performance Edison s testing shows that use of the RtCA has either a positive effect or no effect on performance. In many of our tests, storage system throughput was higher with compression on than without. For example, copying a file compressed with the RtCA from the array back to local storage was 51 percent faster than copying the same file from uncompressed storage. Compression Since use of the RtCA is transparent, the effects of compression are only perceived on the storage array in the form of higher available capacity. The high compression levels possible for active data (as high as 91 percent in Edison s testing) results in lower acquisition costs for storage capacity, while having no effect on performance or user experience. Edison: Array Based Compression Storage Efficiency Technology Page 10

15 Research Results Test Environment The test bed for this study consisted of a VNX 5500 mid-range storage system, the RtCA, a half dozen x86 servers for load generation, and 10GbE and 1GbE networks for NAS and management respectively. Specifically: EMC VNX5500 with GB SAS HDD, 10GbE, iscsi, NAS, and FC-enabled IBM Real-time Compression Appliance model STN6800 with eight 10GbE ports Six IBM System x x3550 Servers with 32GB RAM, three 300GB SAS HDD, SolarFlare SFN5122F 10GbE NICs Blade Network GbE Switches Software used included: Red Hat Enterprise Linux, 6.2, Microsoft Windows Server 2008 R2 Standard Edition, VMware vsphere 4.1, Quest Software Benchmark Factory for Databases, Symantec Net Backup 7 (the leading enterprise class backup and data protection program), IOzone.org benchmark software. 11 SUT Installation The Systems Under Test (SUT) were installed is a data rack and connected to a 1GbE management and a 10GbE data network. Figure 6 presents a logical representation of the configuration. (In reality, the servers, RtCA and VNX array were conventionally mounted in the rack with the switch mounted at the top of the rack.) The servers, were all connected to the 10GbE switch shown here as a single cable for simplicity. There were two 10GbE connections to the storage systems. One connection was made directly to the EMC VNX The second was connected to a port on the RtCA. A second connection was made between the RtCA and the VNX array. Connectivity between the servers and the storage array via the RtCA is completely transparent: the RtCA acts as a bridge. The servers connect to the VNX using the IP address of the array, not to the address of the RtCA port. 11 Edison Group is grateful to Red Hat and Symantec for providing us with the software used in these tests. Edison: Array Based Compression Storage Efficiency Technology Page 11

16 Figure 6: Network Configuration Test Objectives Edison conducted five separate tests with the goal of determining immediate compression rates and compression over time, performance, and transparency. Edison considered two types of transparency in the test environment: configuration and operational transparency. Configuration transparency is determined by whether changes need to be made to server and storage array configurations in order to use the appliance. Operational transparency, an extension of configuration transparency, is determined by whether the appliance is visible to the applications running on the host and, in turn, visible to end users. The tests are described in Table 1 on the following page. Edison: Array Based Compression Storage Efficiency Technology Page 12

17 Table 1: Test Cases Test Name Description 1 File Copy 2 Compression over time 3 IOzone 4 TPC-C 5 Backup to Disk using NDMP Determine system throughput in copying files, and copying files with compression enabled, using both Real-time Compression and VNX Series compression. Demonstrate the difference in capacity utilization of files compressed with RtCA or VNX Deduplication over time after file access/updates. Test maximum filer throughput input and output using IOzone with large files. Compare compression rates in various scenarios, including: OLTP/database use-case Filer load Compression ratio over time Determine the time difference for backing up and restoring data to disk, using the Network Data Management Protocol with and without compression. Test One File Copy Copying files from end-user computers or between servers is one of the most frequently performed computing activities. File Copy Time and Compression Levels Table 2: File Copy to array File Copy No Compression (Copy to VNX Array) Total Time Difference (percent) Total Size on Disk (GB) (EMC VNX) 10:35 min 123 GB Difference (percent) With Compression 9:25 min 13% 11 GB 91% Edison: Array Based Compression Storage Efficiency Technology Page 13

18 Table 3: File Copy from Array File Copy Total Time Difference (percent) Total Size on Disk (GB) (Server Disk) No Compression 12:48 min 120 GB With Compression 6:20 min 51% 120 GB With VNX Decompression 17:10 min -25% 120 GB Copying a collection of files of mixed file sizes and types (see appendices for details) from a server to the array with RtCA, with a compression level of 91 percent compression, took about 13 percent less time than copying without using the RtCA. When copying from the array back to the server, the time differential was increased, with savings of more than 51 percent over uncompressed data copying. System Utilization Results Network Interface Utilization The chart (Figure 8) below shows utilization across the VNX network interface over the test period. As can be seen, there is far greater network utilization for uncompressed data. The average difference is over 200 percent. The limited data points captured by the array exaggerate the peak utilization delta but the average utilization figures were consistent over the test runs. Figure 7: VNX Network Interface Utilization Edison: Array Based Compression Storage Efficiency Technology Page 14

19 CPU Utilization The storage array CPU-utilization results are commensurate with the network utilization levels, though with a less of a performance spike (Figure 9, below).the average utilization difference was about 31 percent. Figure 8: VNX CPU Utilization What the Test Results Mean Copying data from clients to servers or between servers is one of the most common activities on a network. These files comprise a significant portion of the unstructured data that tax storage budgets. Any technology that can make the transfers more efficient and lowers physical capacity requirements will quickly pay for itself. The tests Edison performed were designed to show how the RtCA could improve performance, utilize less storage space and do so with complete transparency. Files were copied to the storage array at the same IP address over the 10GbE network without and the RtCA enabled. This approach illustrated that using RtCA is completely transparent from an end-user perspective. The results show remarkable compression levels 91 percent 12 with a time savings of 13 percent when copying to the array. More importantly, because users perceive the time required for opening of downloading a file differently, when copying files from the array back to the host, the time savings were 51 percent. 12 These compression levels were higher than expected and may be attributable to the nature of the files used. Running the same test with a different mix file sizes and types may have resulted in different compression levels. Edison: Array Based Compression Storage Efficiency Technology Page 15

20 These results demonstrate that the time spent copying files can be significantly reduced with no change in copy practices when the RtCA is used, increasing productivity. The high compression levels point to the savings in physical storage capacity required for file copy services. Test Two Compression over Time By definition, data on a storage system will persist over time. Nevertheless, most data changes, except for archived and other data that, by policy, remains unchanged. A change to data that has been compressed by any of the means evaluated in this study has to have an effect on the space used to store that data. For this study, Edison intended to compare the effect of making changes on capacity utilization over time as a result to changes to large files that had been compressed with IBM RtC and EMC VNX deduplication/compression. We devised several file-copy scenarios to show the effects of file changes on storage system utilization, but due to the differences in the underlying technologies of the two platforms, found that it was impossible to show the differences between the two systems from test data in a meaningful manner. RtCA compresses files as they are created on the array (when copied to the array in this test). When files are later changed and saved, the changed files are also compressed by the RtCA. This results in stable storage capacity utilization over time. As previously described, RtCA leverages variable input to fixed output compression when a file is updated, RtC utilizes its mapping to fill in the potential gaps in the compressed file. It is similar to a journaled file system that keeps track of changes, just inside the compressed file. As a result, files that are changed and saved only grow to the extent that entirely new bits of data are added if the file size grows, the space required to store the file grows in proportion to the degree of file change times the compression ratio. For most files, this means that requirements for capacity growth are nearly flat. When files are compressed utilizing VNX deduplication/compression, the files are saved to the array and then deduplicated/compressed using a post-process. If the files stored are unique, deduplication has no effect; if there are no duplicates, there can be no deduplication. Therefore, file deduplication is only applicable for datastores containing two or more files with nearly identical content. When deduplication is performed, data with unique byte patterns are analyzed and stored, while data that matches already stored data is replaced with a reference that points to the stored bytes. As with Real-time Compression, capacity requirements grow with the increase in unique data in the deduplicated files, also resulting in nearly flat capacity utilization. However, there is an important caveat: VNX deduplication does not occur in real time. The default schedule is to deduplicate files that have not changed over the previous 30 days, though the system can be configured for shorter intervals. This means that accessing previously Edison: Array Based Compression Storage Efficiency Technology Page 16

21 deduplicated files required them to be reconstituted to their original state requiring storage capacity be available for the file and only re-deduplicated at the next scheduled time. This means that though the application of either technology will result in relatively flat storage utilization growth for unchanged files, storage utilization with VNX Deduplication and Compression will result in periods of extreme capacity growth followed by deduplication and lowered capacity utilization. The graph below illustrates the difference between the two platforms. It also shows the tremendous difference between deduplication and Real-time Compression on capacity utilization. For this test, Edison created a test that utilized two identical large 6MB text files, with different names: file1, file2. These files were copied to the array with and without the RtCA and with VNX deduplication enabled. Uncompressed, the files required about 12 MB of space. After a forced VNX deduplication cycle, the files required about 6.1 MB of space. The space required includes the unique data in the files, the metadata describing the differences between the files and any data that has been changed in the files. In comparison, after RtCA compression, the files only required about 800KB of storage space and, when accessed and changed, remained at that size on the array. Figure 9: Illustration of compression over time comparison between EMC VNX Deduplication/Compression versus IBM Real-time Compression. The VNX line shows both the time during which a file has been reconstituted pending the next deduplication cycle and the time between file changes during which the file remains deduplicated. Edison: Array Based Compression Storage Efficiency Technology Page 17

22 What the results mean The results of the compression-over-time testing reveal two important factors when comparing compression technologies. First, EMC is right files that are subject to modification are not suitable for deduplication, requiring more storage capacity be available. For example, log files that are constantly growing and contain data that should be highly deduplicable are not good candidates for VNX Deduplication. These same files will all be transparently compressed by the RtCA at very high compression levels. This means that for organizations with large amounts of slow changing file-based data, adding an RtCA can see significant gains in storage utilization efficiency. Second, when VNX deduplication is applied to occasionally changing files, storage adequate capacity is required to accommodate those files at full size for the period between deduplication cycles. In addition, there may be concerns with fragmentation and other factors that could affect array performance. As a result, storage capacity is not being efficiently utilized and storage system performance could be affected for the worse. Test Three IOzone with Large Files The IOzone benchmark is designed to test the performance of an entire computer system configuration computer, storage and, where applicable, the network connecting storage to the computer. Since one of the goals of this study was to demonstrate what effect, if any, the RtCA had on storage system performance, Edison focused on the network throughput as measured at the storage array. The file-copy tests also show network throughput performance at the array, but not at the sustained levels produced by IOzone. Table 4: IOzone - performance in KB/Sec for all tests with and without compression Comparison with and without Compression Without Compression With Compression Difference Kbytes/Sec Throughput Advantage with Compression Initial write 458, , % Rewrite 466, , % Read 239, , % Re-read 729, , % Random read 548, , % Random write 468, , % Average Compression 2,911, ,524, % Edison: Array Based Compression Storage Efficiency Technology Page 18

23 The table above shows system throughput in KB/sec. for the gamut of IOzone tests with and without Real-time Compression. 13 In all but the case of random writes, performance with the RtCA in use was greater than without the RtCA. The chart, below, showing system throughput, reveals another effect of the RtCA performance is more consistent across the range of tests, indicating that the appliance ameliorates performance bottlenecks inherent on the storage array. Figure 10: System Throughput with IOzone What the Test Results Mean IOzone is a file system benchmark tool, designed to generate and measure a variety of file operations. Though originally intended to evaluate the performance of computer hardware and operating systems, the benchmark tests the performance of the entire environment, making it very useful for testing the performance of network attached file systems under load. The data captured by the benchmark mostly focuses on application performance, since it is application response times that users most care about. Since we were only interested in those aspects of performance related to the network attached storage device and the effects on performance of the RtCA, Edison only utilized the throughput in KB/Sec data generated by the IOzone software. 13 IOzone tests were not run against EMC VNX deduplication/compression. This was because compression occurs at idle times after data is written to the array. The IOzone test bed reads and writes over a period of time, deleting the files created at the end of the test. This means that no array-based compression can occur during the running of the IOzone tests. Therefore, enabling EMC VNX deduplication/compression will have no effect on IOzone test performance. Edison: Array Based Compression Storage Efficiency Technology Page 19

24 The systems were configured so that the tests utilized the same IP address for the storage system, with and without the RtCA compression enabled. This was done in order to demonstrate transparency: no changes to the hosts are required. Overall, the results revealed that file I/O performance was more than 21 percent higher when RtCA compression was enabled, while varying in keeping with the range of operations performed different operations revealed different effects. For example, the peak performance improvement showed most predominantly during the read test, which encompassed copying from the array to the server s local storage. Interestingly, the results for the re-read test showed the smallest advantage, where host OS and array caching can have the greatest effects. It is important to note that implementing the deduplication/compression features native to the VNX5500 array would have no effect on the IOzone tests since deduplication only occurs on unchanged files during idle periods, post-write. From an operations perspective, this illustrates that organizations looking for improvements in storage capacity utilization and storage system throughput cannot get these improvements from their storage array. These organizations can realize throughput improvements by adding spindles to the array, worsening the capacity challenge, or adding Solid State Drive storage which improves throughput, but at a very high price 14. Test Four Database Performance Measured with TPC-C The TPC-C benchmark was designed to test the transaction performance of a complete database server platform hardware and software in transactions per minute. Usually, the results of the benchmark are presented in the context of cost per transaction, a metric that is good for comparing the different platforms. For this study, TPC-C was used solely to generate transaction loads that would demonstrate any effects on performance resulting from using the RtCA. Compression is gaining in importance for database systems as a means of dealing with the exponential growth in the number of transactions and the storage capacity needed to contain that growth. The EMC VNX5500 system used for this study offers file deduplication and compression features, but these features are not suitable for the active data used by a database. In fact, EMC strongly recommends not using deduplication on active database files. Another approach to addressing this challenge is the use of expensive customized hardware platforms with specialized storage subsystems, and changes to the database architecture designed to enable high performance compression. 14 Solid State Drives do improve performance but, by design, they are not really suited for filecopy types of use being better suited for transaction-based applications. Edison: Array Based Compression Storage Efficiency Technology Page 20

25 These systems work quite well, but they require that existing databases be moved to the new systems and database design, as well as the design of the applications using them, be modified to take advantage of the new platform. Figure 11: TPC-C Tests Throughput (TPM-C) Figure 12: TPC-C Tests Response Time Edison s tests show that by adding a RtCA to the NFS storage network not only provides high compression (64 percent in our tests, as high as 90 percent in tests run by IBM in their labs) but also this compression was delivered with no effect on performance (Figure 12). In fact, although the raw TPM-C performance was virtually identical with or without compression, the rate of change for each 50-user increment was higher with the RtCA enabled (Figure 13). This acceleration can be seen in Figure 14 on the next page. Edison: Array Based Compression Storage Efficiency Technology Page 21

26 Figure 13: TPC-C Throughput Trend Comparison What the Test Results Mean The results of Edison s tests show that organizations with large volumes of database transaction data can utilize an RtCA to greatly lower storage capacity requirements with no effect on system throughput. More importantly, there is no need to change database schema or application design in order to accommodate the use of compression. In fact, once the RtCA has been inserted into the storage network, no changes to the environment are required at all. Test Five Backup using NDMP Compression has long been a key feature of data backup systems. Both backup software and the tape drives used for backup have had compression capability for years. Ideally suited to backups, deduplication capabilities can be provided by specially designed deduplicating backup arrays such as HP StoreOnce or EMC Data Domain. Many of the popular backup software products can also provide software-based deduplication capabilities, but doing so requires more powerful backup servers and specially configured disk storage for operation. This study looks at alternative mechanisms for compressing backups on a standard storage array with hardware, rather than backup software. We compare the effects of EMC VNX deduplication/compression on the array with compression provided by the IBM RtCA. Edison: Array Based Compression Storage Efficiency Technology Page 22

27 Table 5: NDMP Backup - Duration with and without compression Policy File Set Backup Without Compression File Set Backup With RtCA Compression File Set Backup With VNX Compression File Set Restore Without Compression File Set Restore With RtCA Compression File Set Restore With VNX Compression Original File Size MB Compressed File Size Duration (Minutes) File Size Advantage Duration Advantage N/A 32 N/A N/A % Over Uncompressed 21% Over VNX Compression 43% Over Uncompressed 53% Over Uncompressed 55% Over VNX Compression Since deduplication is performed as a separate process, it has no effect on backup write time N/A 48 N/A N/A N/A 27 N/A N/A 52 N/A 44% over uncompressed 48% Advantage over VNX Deduplication 8% Slower than Uncompressed 93% slower than RtC Compression The results depicted in the chart above show that the compression efficiency 73 percent compression was so much greater with IBM Real-time Compression than with either no compression, or with EMC VNX compression. The time required to backup files with RtCA enabled required was 53 percent less time than the time required for uncompressed data, and 55 percent less than data that had been deduplicated via VNX deduplication. Edison: Array Based Compression Storage Efficiency Technology Page 23

28 Restoring files via RtCA clearly shows the power of the platform. Restoration of RtCA compressed files required 44 percent less time than the time required for uncompressed files and 48 percent less time than the time required for restoring deduplicated files. Overall, using VNX deduplication did improve capacity utilization efficiency requiring 23 percent less space than for non-deduplicated data but the need to reconstitute the files for restoration apparently had a slightly negative performance impact of 8 percent. What the Test Results Mean Mostly for reasons of performance, for many organizations, backup to disk (also called Disk-to-Disk or D2D) has increasingly become the preferred approach to data protection, relegating tape to archiving and other off-line data projection scenarios. 15 However, disk is still more expensive than tape, especially in enterprise storage arrays. Compression has long been used to reduce the amount of storage media required for backup operations, so using compression for D2D backup is a logical extension of current practices. These tests show that backing up data with an RtCA as part of the configuration can deliver twice the performance possible when backing up directly to a Virtual Tape Library. These savings can have a significant effect on the duration of backup windows, lowering the operational impact of backups on active systems. When restoring compressed files, RtCA required nearly half the time than for uncompressed or deduplicated data. A recent report 16 revealed that the cost of downtime could be as high as $5,000 per minute or $300,000 per hour for some organization. For these organizations, savings of nearly 50 percent in the time required file restoration could pay for the cost of a Real-time Compression Appliance. 15 How Disk-to-Disk Backup Enables Faster, More Cost-Effective Backups, Restores, and Nearline Storage, Adaptec, by PMC. 16 Understanding the Cost of Data Center Downtime An Analysis of the Financial Impact on Infrastructure Vulnerability from Emerson Network Power. ( Edison: Array Based Compression Storage Efficiency Technology Page 24

29 Conclusion Based on our research, Edison found that utilizing Real-time Compression in a primary storage environment results in overall capacity savings. Because there is less data to store, there is less capacity required. This results in both upfront savings on storage and downstream savings on capital expenditure and energy costs lower storage requirements means less management, a smaller infrastructure footprint which in turn requires less energy to power. Edison found utilizing IBM RtCA resulted in IT resource savings as well, given that less time required to back up more data saves an IT administrator s time. For example, in its NDMP test Edison found that backup via the RtCA over the NDMP protocol using Symantec NetBackup required 53 percent less time than backing up without compression, and 76 percent less time than backing up with VNX deduplication/compression. Edison also validated IBM s claims in the areas of transparency, performance and maximum compression over time. Research findings include: Complete transparency: The RtCA is invisible to the server operating systems and applications, which results in no administrative overhead to use RtCA after the system has been configured. EMC VNX Series, in contrast, is designed to compress inactive content and is not suitable for active data. Therefore, though the EMC VNX demonstrates file copy and backup transparency, native EMC compression is neither recommended nor practical for highly active data utilization in applications such as databases. No impact on performance: Edison found no impact on performance with RtCA online and in fact found an improvement in areas such as backup and IOzone, where system throughput was almost 57 percent greater for read operations with the RtCA compression enabled than without compression. Compressing Oracle database files with the RtCA resulted in no performance differences, while the EMC VNX does not recommend using its compression features with active databases. IBM Real-time Compression, on the other hand, works well with both structured (database) and unstructured (file-based) data, utilized in both physical and virtual operating platforms. That said, because IBM Real-time Compression is utilized on active, primary data, there are some specific data types that are ideally suited for real-time compression. Those data types that will see the highest compression rates include: Text, database, CAD/CAM, VMware virtual environments, Office files, and raw video. Edison: Array Based Compression Storage Efficiency Technology Page 25

30 Because IBM Real-time Compression is best suited for the data types listed above, in turn it is best utilized in certain vertical industries that rely heavily on utilizing specific applications, such as CAD/CAM or Microsoft Office, or on raw video footage. Those industries that will see the biggest benefit from Real-time Compression include: retail and manufacturing design, engineering, up-stream oil and gas, pre-production video, Telco, life sciences, and insurance. In conclusion, the test results in this white paper prove that using IBM Real-time Compression is transparent at installation and in use and that there is no degradation to average performance with compression turned on. Significantly, in several of our test scenarios, system throughput was greater with RtCA enabled than without its use. Finally, Edison found the IBM RtCA delivers superior file system compression over its competitor in primary data environments. Edison: Array Based Compression Storage Efficiency Technology Page 26

31 Appendices Test Setup To perform its tests Edison set up a lab environment consisting of the following hardware, software and key system configurations features: Servers: IBM System x x3550 w/ 32GB RAM, Three 300GB SAS Drives, SolarFlare SFN5122F 10Gb Ethernet NICs Software: Red Hat Enterprise Linux 6.2 Microsoft Windows Server 2008 R2 Standard Edition VMware vsphere 4.1 Quest Software Benchmark Factory for Databases to run TPC/C benchmark tests Symantec Net Backup 7 IOzone File System Benchmark Storage: EMC VNX5500, GB Drives, 10GbE, iscsi, FC IBM Real-time Compression Appliance STN6800 with eight 10GbE ports (other features standard for the 6800 model) 17 Network: IBM Blade Network GbE Switch 17 There are two models of IBM Real-time Appliance. The IBM STN6500 Real-time Appliance has sixteen 1GB Ethernet ports, which supports eight connections between NAS systems and network switches. The STN6800 model supports multiple 1GbE and 10GbE NAS connections. Edison: Array Based Compression Storage Efficiency Technology Page 27

32 A Note on the Use of Benchmarks Edison made use of IOzone and TPC-C tests in order to generate workloads to read and write data to the storage array. The software was not used to achieve maximum performance for the systems under test or create a publishable benchmark result. The goal was to attain reasonable performance levels that would be consistent across the test runs. Test Results In the course of this validation, Edison ran its tests utilizing an EMC VNX Series storage system and x86 servers running Windows Server 2008 R2 and Red Hat Enterprise Linux. In addition, comparisons were made between the performance of the IBM Real-time Compression Appliance and NAS-based compression technologies. Test Case 1 File Copy The tests were run with the following parameters: IOzone was used to create two 400 MB and two 144 MB files. A single folder contained 16 files, totaling six GB of data from MB files and one 144 MB file. Each run copied 120 GB of data in 16 folders in a target directory using 20 simultaneously running scripts utilizing a CPU affinity parameter for targeting single thread/core of the CPUs. Edison is not reporting results from testing EMC VNX Deduplication/Compression for two reasons: 1. Different architectures: EMC VNX deduplication/compression occurs when the system is idle, and then only on files that have not had subsequent changes over a preset period. This means that there is no difference in time required for copying files to the array. Therefore, there was no reason to compare server to array copy times. 2. Inconsistent results: Compression ratios after deduplication/compression were inconsistent across several tests. In keeping with the design of the EMC VNX, it is possible that keeping the system idle for longer periods would have resulted in greater and more consistent compression. However, doing so was not practical within the scope of this study. In addition, the time required for copying VNX deduplicated files from the array back to the server was also inconsistent, probably due to the differences in compression levels for the individual files. There was some Edison: Array Based Compression Storage Efficiency Technology Page 28