http://www.linux-mag.com/cache/7497/1.html Metadata Performance of Four Linux File Systems Jeffrey B. Layton Wed, 2 Sep 2009 Introduction Using the principles of good benchmarking, we explore the metadata performance of four linux file systems using a simple benchmark, fdtree. In a previous article, the case was made for how low file system benchmarks have fallen. Benchmarks have become the tool of marketing to the point where they are mere numbers and do not prove of much use. The article reviewed a paper that examined nine years of storage and file system benchmarking and made some excellent observations. The paper also made some recommendations about how to improve benchmarks. This article isn't so much about benchmarks as a product, but rather it is an exploration looking for interesting observations or trends or the lack thereof. In particular this article examines the metadata performance of several Linux file systems using a specific micro-benchmark. Fundamentally this article is really an exploration to understand if there is any metadata performance differences between 4 Linux file systems (ext3, ext4, btrfs, and nilfs) using a metadata benchmark called fdtree. So now it's time to eat our dog food and do benchmarking with the recommendations previously mentioned. Start at the Beginning - Why? The previous article made several observations about benchmarking, one of which is that storage and file system benchmarks seldom, if ever, explain why they are performing a benchmark. This is a point that is not to be underestimated. Specifically, if the reason why the benchmark was performed can not be adequately explained, then the benchmark itself becomes suspect (it may just be pure marketing material). Given this point, the reason the benchmark in this article is being performed is to examine or explore if, and possibly how much, difference there is between the metadata performance of four Linux file systems using a single metadata benchmark. The search is not to find which file system is the best because it is a single benchmark, fdtree. Rather it is to search for differences and contrast the metadata performance of the file systems. Why is examining the metadata performance a worthwhile exploration? Glad that you asked. There are a number of applications, workloads, and classes of applications that are metadata intensive. Mail servers can be very metadata intensive applications because of the need to read and write very small files. Sometimes databases have workloads that do a great deal of reading and writing small files. In the world of technical computing, many bioinformatic applications such as gene sequencing applications, do a great deal of small reads and writes. The metadata benchmark used in this article is called fdtree. It is a simple bash script that stresses the metadata aspects of the file system using standard *nix commands. While it is not the most well-known benchmark in the storage and file system world, it is a bit better known in the HPC (High Performance Computing) world. An Examination of fdtree Before jumping into the results, it is appropriate and highly recommended to examine the benchmark itself. fdtree is a simple bash script that performs four different metadata tests: * Directory creation * File creation * File removal * Directory Removal It creates a specified number of files of a given size (in blocks) in a top-level directory. Then it creates a specified number of subdirectories and then in turn subdirectories are recursively created up to a specified number of levels and are populated with files. Directory Creation This phase of the benchmark begins by creating the number of specified directories in the main directory using the simple "mkdir" command in a bash function "create_dir". mkdir $base_name"L"$nl"D"$nd"/" The bash variables specify the details of the directory names. The next step is to call the "create_dir" function recursively with a different "base name" (directory) to create all of the required directories. create_dirs $((nl-1)) $base_name"L"$nl"D"$nd"/" File Creation This step of the benchmark creates the required number of files using the "dd" command in a bash function, "create_files". dd if=/dev/zero bs=4096 count=$fsize of=$file_name > /dev/null 2>&1 To create files in the subdirectories, the bash function is called recursively. As part of the benchmark, the number of 4 KiB blocks is specified ($fsize). File Removal The third function in the benchmark is to remove the files that were created. This is done with the standard "rm" command in a function called "remove_files". rm -f $file_name The function "remove_files" is called recursively to remove all of the files. Directory Removal The fourth and final function in the benchmark is to remove the directories. This is done in a bash function "remove_dirs" using the *nix command "rmdir $dir_names" rmdir $dir_names The function "remove_dirs" is called recursively to remove all of the directories. Overall the script uses standard *nix commands for the benchmark. It does not use any recursive options for any of the *nix commands. It stresses the metadata capabilities of the file system because of the potentially large number of files and directories. The one interesting thing that the test does, is round the results, time and rates, to integer values. So there could be times when the time to execute the test could be 0 seconds. That is, the test ran in less than 1 second. Running the benchmark In the benchmark exploration in this article, fdtree was used in 4 different approaches to stressing the metadata capability: * Small files (4 KiB) + Shallow directory structure + Deep directory structure * Larger files (4 MiB) + Shallow directory structure + Deep directory structure The two file sizes, 4 KiB (1 block) and 4 MiB (1,000 blocks) were used to get some feel for a range of performance as a function of the amount of data. The two directory structures were used to stress the metadata in different ways to discover if there is any impact on the metadata performance. The shallow directory structure means that there are many directories but not very many levels down. The deep directory structure means that there are not many directories at a particular level but that there are many levels. To create the specific parameters for fdtree used in the exploration, there were three overall goals: * Keep the total run time to approximately 10-12 minutes at a maximum * Keep the total data for the two directory structures approximately the same * Keep the run time for each of the four functions greater than 1 minute if possible All four functions were not always run for 1 minute, sometimes only for a few seconds. These will be noted in the results. Next we show the command lines for the four combinations. Small Files - Shallow Directory Structure ./fdtree.bash -d 20 -f 40 -s 1 -l 3 This command creates 20 sub-directories from each upper level directory at each level ("-d 20") and there are 3 levels ("-l 3"). It's a basic tree structure. This is a total of 8,421 directories. In each directory there are 40 files ("-f 40") each sized at 1 block (4 KiB) denoted by "-s 1". This is a total of 336,840 files and 1,347,360 KiB total data. Small Files - Deep Directory Structure ./fdtree.bash -d 3 -f 4 -s 1 -l 10 This command creates 3 sub-directories from each upper level directory at each level ("-d 3") and there are 10 levels ("-l 10"). This is a total of 88,573 directories. In each directory there are 4 files each sized at 1 block (4 KiB). This is a total of 354,292 files and 1,417,168 KiB total data. Medium Files - Shallow Directory Structure ./fdtree.bash -d 17 -f 10 -s 1000 -l 2 This command creates 17 sub-directories from each upper level directory at each level ("-d 17") and there are 2 levels ("-l 2"). This is a total of 307 directories. In each directory there are 10 files each sized at 1,000 blocks (4 MiB). This is a total of 3,070 files and 12,280,000 KiB total data. Medium Files - Deep Directory Structure ./fdtree.bash -d 2 -f 2 -s 1000 -l 10 This command creates 2 sub-directories from each upper level directory at each level ("-d 2") and there are 10 levels ("-l 10"). This is a total of 2,047 directories. In each directory there are 2 files each sized at 1,000 blocks (4 MiB). This is a total of 4,094 files and 16,376,000 KiB total data. Each test was run 10 times with the four combinations of the file systems (ext3, ext4, btrfs, nilfs). The test system used for these tests was a stock CentOS 5.3 distribution but with a 2.6.30 kernel and e3fsprogs was upgraded to the latest version as of the writing of this article, 1.41.9. The tests were run on the following system: * GigaByte MAA78GM-US2H motherboard * An AMD Phenom II X4 920 CPU * 8GB of memory * Linux 2.6.30 kernel * The OS and boot drive are on an IBM DTLA-307020 (20GB drive at Ulta ATA/100) * /home is on a Seagate ST1360827AS * There are two drives for testing. They are Seagate ST3500641AS-RK with 16 MB cache each. These are /dev/sdb and /dev/sdc. Only the first Seagate drive was used, /dev/sdb for all of the tests. For all 4 file systems, the defaults were used in building the file systems. For btrfs, version btrfs-progs v0.18. For nilfs2, nilfs-utils-2.0.14 was used. Both ext3 and ext4 were mounted with "data=ordered" since this is recommended practice to prevent data loss. Benchmark Results This section presents the results of the testing (exploration). The results are presented in tables listing the average value and just below it, in red, is the standard deviation. This is done for each of the four combinations and each of the four file systems. The first combination tested was for small files (4 KiB) with a shallow directory structure. Table 1 below lists the results with an average value and just below it, in red, is the standard deviation. Table 1 - Benchmark Times Small Files (4 KiB) - Shallow Directory Structure File System Directory Create File Create File Remove Directory Remove (secs.) (secs.) (secs.) (secs.) ext3 13.00 342.90 69.40 1.30 3.61 42.69 6.92 0.46 ext4 10.60 327.20 58.10 1.40 0.92 4.89 1.87 0.92 btrfs 8.80 335.00 65.30 1.40 0.40 1.00 0.78 0.66 nilfs2 9.10 345.70 51.60 1.20 0.30 8.14 0.92 0.40 The first test, directory creates, had an average run time of 12 seconds for all four file systems, so the results may not be that meaningful. In addition, the directory remove test ran in about 1 second or less. Consequently, this test may not have much value. Table 2 below lists the performance results with an average value and just below it, in red, is the standard deviation. Table 2 - Performance Results of Small Files (4 KiB) - Shallow Directory Structure Directory File File File Directory File Create Create Create Remove Remove System (Dirs/sec) (Files/ (KiB/sec) (Files/ (Dirs/sec) sec) sec) ext3 695.20 993.70 3,975.90 4,900.30 7,578.80 177.37 94.66 378.91 473.28 1,684.40 ext4 800.00 1,029.10 4,118.30 5,803.40 7,368.30 69.88 15.21 60.59 111.90 2,157.37 btrfs 958.40 1,005.00 4,021.70 5,167.10 7,017.40 46.80 3.00 12.01 78.22 2,174.42 nilfs2 925.70 974.70 3,889.20 6,529.20 7,578.80 27.90 21.67 88.54 112.75 1,684.40 The second combination tested was for small files (4 KiB) with a deep directory structure. Table 3 below lists the benchmark times with an average value and just below it, in red, is the standard deviation. Table 3 - Benchmark Times Small Files (4 KiB) - Deep Directory Structure File System Directory Create File Create File Remove Directory Remove (secs.) (secs.) (secs.) (secs.) ext3 46.20 182.40 53.70 14.60 26.97 72.55 24.78 7.55 ext4 187.00 443.20 192.50 73.30 11.22 7.69 12.51 42.09 btrfs 102.40 398.6 132.50 38.10 0.66 1.91 0.67 0.70 nilfs2 108.20 417.30 122.10 37.20 2.68 6.48 3.39 0.60 For these tests, the first test, directory creates, took about 40 seconds for ext3 (the fastest). This time is fairly small and, consequently, the results may not be as applicable the other tests which had a much longer run time. The last test, directory removes, took 11 seconds for ext3 (the fastest). Again, this time is fairly quick so it may not be as useful because the time is so short. Table 4 below lists the performance results with an average value and just below it, in red, is the standard deviation. Table 4 - Performance Results of Small Files (4 KiB) - Deep Directory Structure Directory File File File Directory File Create Create Create Remove Remove System (Dirs/sec) (Files/ (KiB/sec) (Files/ (Dirs/sec) sec) sec) ext3 783.90 927.90 3,713.00 3,180.70 2,452.40 39.08 16.58 65.88 209.90 207.90 ext4 475.00 799.10 3,198.00 1,848.00 1,539.60 29.05 13.45 53.73 124.31 201.76 btrfs 864.30 888.10 3,554.80 2,673.60 2,324.90 5.76 4.23 16.92 13.87 42.57 nilfs2 818.60 848.50 3,396.40 2,903.60 2,380.80 19.11 12.71 51.73 75.52 36.60 The third combination tested was for medium files (4 MiB) with a shallow directory structure. Table 5 below lists the benchmark times with an average value and just below it, in red, is the standard deviation. Table 5 - Benchmark Times Medium Files (4 MiB) - Shallow Directory Structure File System Directory Create File Create File Remove Directory Remove (secs.) (secs.) (secs.) (secs.) ext3 0.30 174.90 17.40 0.00 0.46 17.46 3.47 0.00 ext4 0.20 156.80 11.80 0.20 0.40 4.75 2.99 0.40 btrfs 0.50 114.40 15.60 0.10 0.50 1.11 0.49 0.30 nilfs2 0.70 196.30 7.50 0.20 0.78 3.07 2.87 0.40 For these tests, the first test, directory creates, took less than 1 second. This time is very small and, consequently, the results are not as applicable as some of the other tests. The file removes test took about 10-15 seconds. Again this is a very short time and the results may not be as applicable. The last test, directory removes, took 0-1.4 seconds. This time too, is very short. Table 6 below lists the performance results with an average value and just below it, in red, is the standard deviation. Table 6 - Performance Results of Medium Files (4 MiB) - Shallow Directory Structure Directory File File File Directory File Create Create Create Remove Remove System (Dirs/sec) (Files/ (KiB/sec) (Files/ (Dirs/sec) sec) sec) ext3 92.10 17.30 70,889.80 182.30 0.00 140.69 1.90 6,798.06 32.53 0.00 ext4 61.40 18.90 78,393.20 278.30 61.40 122.80 0.54 2,252.90 75.69 122.80 btrfs 153.50 26.20 107,342.50 196.20 30.70 153.50 0.60 1,063.70 6.37 92.10 nilfs2 122.70 15.00 62,572.00 442.50 61.40 133.80 0.00 968.91 90.62 122.80 The fourth and final combination tested was for medium files (4 MiB) with a deep directory structure. Table 7 below lists the benchmark times with an average value and just below it, in red, is the standard deviation. Table 7 - Benchmark Times Medium Files (4 MiB) - Deep Directory Structure File System Directory Create File Create File Remove Directory Remove (secs.) (secs.) (secs.) (secs.) ext3 2.70 248.30 18.80 1.80 0.78 9.99 4.07 1.08 ext4 3.20 219.50 13.40 1.20 0.75 1.12 4.72 0.40 btrfs 2.40 159.30 16.20 1.10 0.49 1.42 1.17 0.30 nilfs2 2.50 287.70 11.50 1.40 0.50 10.67 0.50 0.49 The first test, directory creates, took 2-3 seconds, which is very short. The time for the third test, file removal, was also fairly short at 11-19 seconds. The last test, directory removes, was extremely fast at less than 2 seconds. These three results are somewhat suspect because of short run time. Table 8 below lists the performance results with an average value and just below it, in red, is the standard deviation. Table 8 - Results of Medium Files (4 MiB) - Deep Directory Structure Directory File File File Directory File Create Create Create Remove Remove System (Dirs/sec) (Files/ (KiB/sec) (Files/ (Dirs/sec) sec) sec) ext3 818.30 16.20 66,053.10 225.60 1,518.00 213.10 0.60 2,515.72 35.42 658.48 ext4 671.70 18.10 74,607.50 331.50 1,842.20 147.98 0.30 380.54 112.06 409.60 btrfs 886.60 25.20 102,807.40 253.20 1,944.60 167.06 0.40 917.56 17.72 307.20 nilfs2 852.50 13.70 56,998.60 356.50 1,637.40 170.50 0.64 2,122,26 15.50 501.66 Discussion of Results There are four different combinations that were tested for the four file systems. Comparing the file systems is interesting but also comparing the same file system for the different tests is interesting as well. Firstly, let's examine the shallow directory structure results (Tables 2 and 6). For small files (4 KiB) the four file systems performed about the same for the file create and file removal tests (the directory create and remove tests ran too quickly to be really useful). All four file systems achieved about 1,000 file creates per second or about 4,000 KiB per second (see Table 2). But medium files (4 MiB) produced very different results. For this case, btrfs was almost twice as fast as nilfs and 50% faster than ext3 or ext3 with respect to the number of files creates per second or throughput (KiB/s) (see Table 6). However, the small file case produced 109X more files, 27X more directories, but only produced about 1/10-th the total amount of data. This points out the extreme pressure that is put on metadata performance of file systems for small files. Second, we can perform the same comparison for deep directory structure (Tables 4 and 8). Small files (1 KiB) put extreme pressure on file system metadata performance as with the shallow directory structure. Examining the results, the following observations are made: * Small files put extreme pressure on metadata performance regardless of file systems * For small files, a shallow or deep directory structure did not appreciably impact metadata performance * For larger files, a shallow or deep directory structure also did not appreciably impact metadata performance * For small files, btrfs has good file creation performance but file removal performance is not as good as ext3 and ext4 at this time * For larger files, btrfs has both excellent file creation and removal performance relative to the other 3 file systems * Log-base file systems such as nilfs2 should work well with metadata tests. But the developers are evolving the garbage collection (gc) algorithm which should improve performance. * The standard deviation, or the spread in the data was much greater for ext3 than the other three file systems. The reason(s) for this are not known. This is the first attempt at useful benchmarks and analysis for Linux file systems using the approach outlined in a previous article around benchmarking. There will be future articles that use the same basic tenants. But be ready because this approach introduces a great deal of data into the article. But overall it does give much more information that a quick table and a conclusion "file system X is better" (usually followed by a run for cover). Please let me know in the forums if this type of article is useful.