Introduction: The Flexible Filesystem Benchmark (FFSB) is a filesystem performance measurement tool. It is a multi-threaded application (using pthreads), written entirely in C with cross-platform portability in mind. It differs from other filesystem benchmarks in that the user may supply a profile to create custom workloads, while most other filesystem benchmarks use a fixed set of workloads. As of version 5.1, it supports seven different basic operations, support for multiple groups of threads with different operation mixtures, support for operation across multiple filesystems, and support for filesystem aging prior to benchmarking. Differences from version 4.0 and older: Version 5.0 and above represent almost a total re-write and many things have changed. In version 5.0 and above FFSB moved to a time-regulated run versus doing a set number of different operations and timing the whole thing. This is primarily to better deal with the use of multiple threadgroups which would otherwise not be synchronized at termination time. Additionally, the FFSB configuration file format has changed in version 5.0, although we do support old-style configuration files along with a run-time passed on the command line. In this mode, version 5.0 and above ignores the iterations parameter, and simply uses the time specified on the command line. Behaviorally, most of the old operations are the same -- sequential reads and sequential writes work as they did before. One change in version 5.0 is the skip-read behavior of reading then seeking forward a fixed amount then reading again is removed, we now support fully randomized reads and writes from random offsets within the file. Version 4.0 didn't support overwrites (only appends) so we interpret writes in old config files to be append operations. On Linux, CPU utilization information will only be accurate for systems using NPTL, older Linuxthreads systems will probably only see zeros for CPU utilization because Linuxthreads is non-compliant to POSIX. Version 4.0 and older could be recompiled to work on Linuxthreads, but in 5.0 and later we no longer support this. We no longer support the "outputfile" on the command line. One should simply use tee or similar to capture the output. FFSB unbuffers standard out for this purpose, and errors are sent on standard error. Global options: There are eight valid global options placed at the beginning of the profile. Three of them are required: num_filesystems (number of filesystems), num_threadgroups (number of threadgroups), and time (running time of the benchmark). The other five options are: directio - each call to open will be made using O_DIRECT alignio - aligns all block operations for random reads and writes on 4k boundaries. bufferedio - currently ignored: it is intended to use libc fread,rwrite, instead of just unix read and write calls verbose - currently ignored callout - calls and external command and waits for its termination before FFSB begins the benchmark phase. This is useful for synchronizing distributed clients, starting profilers, etc. They must be specified in the above order (num_filesystems, num_threadgroups, time, directio, alignio, bufferedio, verbose, callout). Filesystems: Filesystems are specified to FFSB in the form of a directory. FFSB assumes that the filesystem is mounted at this directory and will not do any verification of this fact beyond ensuring it can read/write to the location. So be careful to ensure something with enough space to handle the dataset is in fact mounted at the specified location. In the filesystem clause of the profile, one may set the starting number of files and directories as well as a minimum and maximum filesize for the filesystem. One may also specify the blocksize used for creating the files separately in the filesystem clause. Also, if a filesystem is to be aged, a special threadgroup clause may be embedded in a filesystem clause to specify the operation mixture and number of threads used to age the filesystem. This threadgroup is run until filesystem utilization reaches the specified amount. Inheritance -- if you are using multiple filesystems, all attributes except the location should be inherited from the previous filesystem. This is done to make it easier to add groups of similar filesystems. In this case, only the location is required in the filesystem clause. As of version 5.1, filesystem re-use is supported if a given filesystem hasn't been modified beyond it's orginal specifications (number of files and directories is correct, and file sizes are within specifications). This can be a huge time saver if one wishes to do multiple runs on the same data-set without altering it during a run, because the fileset doesn't need to be recreated before each run. To do this, specify "reuse=1" in the filesystem clause, and FFSB will verify the fileset first, and if it checks out it will use it. Otherwise, it will remove everything and re-create the filesets for that filesystem. Threadgroups: An arbitrary number of threadgroups with differing numbers of threads and operation mixes can be specified. The operations are specified using a weighting for each operation, if an operation isn't specified it's weighting is assumed to be zero (not used). "Think-time" for a threadgroup may also be specified in millisecond amounts using the "op_delay" parameter, where every thread will wait for the specified amount between each operation. Operations: All operations begin by randomly selecting a filesystem from the list of filesystems specified in the profile. The distribution aims to be uniform across all filesystems. The seven operations are: reads - read() calls with an overall amount and a blocksize operates on existing files. Care must be taken to ensure that the read amount is smaller than the size of any possible file. If random_read is specified, then the each individual blocks will be read starting from a random point with the file, and this will continue until the entire amount specified has been read. This offset of each random block will be totally random to the byte level, unless the "alignio" global parameter is on, and then the reads will be 4096 byte aligned. This is generally recommended. readall - Very similar to read above, except it doesn't take an amount; it simply reads the entire file sequentially using the read_blocksize. This is useful for situations where different filesystems have differently sized files, and sequential read patterns across all filesystems are desired. writes - write() calls with an overall amount and blocksize this is an overwrite operation and will not enlarge an existing file, again one must be careful not to specify a write amount that is larger than any possible file in the data set. If random_write is specified, then the each individual blocks will be written starting from a random point with the file, and this will continue until the entire amount specified has been written out. This offset of each random block will be totally random to the byte level, unless the "alignio" global parameter is on, and then the writes will be 4096 byte aligned. This is generally recommended. If the fsync_flag parameter for the threadgroup is non-zero, then after all of the write calls are finished, fsync() will be called on the file descriptor before the file is closed. creates - creates a file using open() call and determines the size randomly between on the constraints (min_filesize and max_filesize) for the selected filesystem. Write operations will be done using the same blocksize as is specified for the write operation. deletes - calls unlink() on a filename and removes it from the internal data-structures. One must be careful to ensure there are enough files to delete at all times or else the benchmark will terminate. appends - calls write() using the append flag with an overall amount and a blocksize to be appended onto a randomly chosen file. metas - this is actually a mix of several different directory operations. Each "meta" operation consists of two directory creates, one directory remove, and a directory rename. These operations are all carried out separately from the other 5 operations. Operation accounting: Each operation which uses a blocksize counts each read/write of a blocksize as an operation (reads,writes,creates, and appends) whereas deletes and metas are considered single operations. Running the benchmark: There are three phases to running the benchmark, aging, fileset creates, and the benchmark phase. The create phase is carried out across all filesystems simultaneously with one dedicated thread per filesystem. After the create phase, sync() is called to ensure all dirty data gets written out before the benchmark phase begins, and sync() is again called at the end of the benchmark phase. The time in sync() at the end of the benchmark phase is counted as part of the benchmark phase. Caveats/Holes/Bugs: Aging and aging across multiple filesystems simultaneously hasn't been tested very much. If *any* i/o operation or system call/libc call fails, the benchmark will terminate immediately. The parser doesn't handle mal-formed or incorrect profiles very well (or at all). The parser doesn't check to make sure all of the appropriate options have been specified. For example, if writes are specified in a threadgroup but write_blocksize isn't specified, the parse won't catch it, but the benchmark run will fail later on. Configuration Files (new style): New Style Configuration allows for arbitrary newlines between lines, and comments using '#' at the start of a line. Also it allows tabs, whitespace before and after configuration parameters. The new style configuration file is broken up into three main parts: global parameters, filesystems, and threadgroups The sections must be in the above order. Global parameters: Global parameters are described above, the first three are always required. Example: ---------- num_filesystems=1 num_threadgroups=1 time=30 # time is in seconds directio=0 # don't use direct io alignio=1 # align random IOs to 4k bufferedio=0 # this does nothing right now verbose=0 # this does nothing right now # calls and external command and waits # everything until the newline is taken # so you can have abritrary parmeters callout=synchronize.sh myhostname --------- All of these must appear in this order, though you can leave out the optional ones. Filesystems: Filesystems describe different logical sets of files residing in different directorys. There is no strict requirement that they actually be on different filesystems, only that the directory specified already exists. Filesystems are specified by a clause with a filesystem number like this: [filesystem0] location=/mnt/testing/ num_files=10 num_dirs=1 max_filesize=4096 min_filesize=4096 [end0] The clause must always begin with [filesystemX] and end with [endX] where X is the number of that filesystem. You should start wiht X = 0, and increment by one for each following filesystem. If they are out of order, things will likely break. The required information for each filesystem is: location, num_files, num_dirs, max_filesize, and min_filesize. Beyond those the following four options are supported: reuse=1 # check the filesystem to see if it is reusable # filesystem aging, three components required # takes agefs=1 to turn it on # then a valid threadgroup specification # then a desired utilization percentage agefs=1 # age the filesystem according to the following threadgroup [threadgroup0] num_threads=10 write_size=40960 write_blocksize=4096 create_weight=10 append_weight=10 delete_weight=1 [end0] desired_util=0.20 # In this case, age until the fs is 20% full create_blocksize=4096 # specify the blocksize to write() # for creating the fileset, defaults to 4096 age_blocksize=4096 # specify the blocksize to write() for aging Also, to allow lazy people to use lots of filesystems, we support filesystem inheritance, which simply copies all options but the location from the previous filesystem clause if nothing is specified. Obviously, this doesn't work for filesystem0. (May not work for aging either?) Full blown filesystem clause example: ---- [filesystem0] # required parts location=/home/sonny/tmp num_files=100 num_dirs=100 max_filesize=65536 min_filesize=4096 # aging part agefs=0 [threadgroup0] num_threads=10 write_size=40960 write_blocksize=4096 create_weight=10 append_weight=10 delete_weight=1 [end0] desired_util=0.02 # age until 2% full # other optional commands create_blocksize=1024 # use a small create blocksize age_blocksize=1024 # and smaller age create blocksize reuse=0 # don't reuse it [end0] -- Threadgroups: Threadgropus are very similar to filesystems in that any number of them can be specified in clauses, and they must be in order starting with threadgroup0. Example: --- [threadgroup0] num_threads=32 read_weight=4 append_weight=1 write_size=4096 write_blocksize=4096 read_size=4096 read_blocksize=4096 [end0] --- In a threadgroup clause, num_threads is required and must be at least 1. Then, at least one operation must be given a weight greater than 0 to be a valid threadgroup. Operations can be given a weighting of 0, and in this case they are ignored. Certain operations will also require other commands, for example, if read_weight is greater than zero, then one must also include a read_size and a read_blocksize. Here's the table of requirements and options: Operation Requirements Options -- -- -- read_weight read_size, read_blocksize read_random readall_weight read_blocksize none write_weight write_size, write_blocksize write_random,fsync_file create_weight write_blocksize or create_blocksize none append_weight write_blocksize, write_size none delete_weight none none meta_weight none none Other threadgroup options: op_delay=10 # specify a wait between operations in milli-seconds bindfs=3 # This allows you to restrict a threadgroup's operation # to a specific filesystem number. Currently only # binding to one specific filesystem is supported