changes.txt

硬盘速度测试 linux下 源码 c/c

V1.0 (capps):
Capps:
	Beginning of the code base.
Isom:
	Added reread
	Added rewrite
	Added read backwards
	Added lseek+read
	Added lseek+reread

Capps:
	Added more accurate time collection method.
	Added alignment in the on chip Cache code.
	Added change step when passing 16 Meg size file.

Capps:
	Added auto+ to purge on chip cache.

kcollins:
	replaced the lseek+read &reread test with random reads and writes

Capps:
	Replaced reverse re-read with record rewrite. This gives
	both source and destination on chip cache hits.

Capps:
	added auto+multi
	Support for multiple buffers in the iozone.

Capps:
	Removed the recursion through main().
	Cleaned up the printout when not in auto mode.
	Added support for stride-read. ( Manual mode only )

Capps:
	Cleanup so it will build for bsd4_2 ( C series machines )

Capps:
	Cleanup on frontend. Now uses getopt() and has a real parser.
	Cleanup on error handling.
	Added throughput tests.
	Eliminated page faults in the throughput tests.

Capps:
	Made all perf tests table driven.
	Add back Bill's Copyright.
	ansify, prototypes, scope limitations.

V2.1 (kcollins):
	Simplified auto mode logic.  auto test now runs from
	MEGABYTES_START to MEGABYTES_END and from RECLEN_START
	to RECLEN_END with values multiplied by MULTIPLIER 
	each iteration.  Range set to (4K...16M) for RECLEN
	and (1M...512M) for MEGABYTES.  (the physical I/O
	tests for RECLEN <4K take several hours on a 1200).

	Enlarged MAXBUFFERSIZE to 16MB for large record tests.

	Added error checking for mallocs (!-).

	Changed multibuffer code to use all of MAXBUFFERSIZE
	as a circular buffer, with the number of buffers
	varying as MAXBUFFERSIZE/reclen.  This corrects 
	problems where MAXBUFFERSIZE*MAXBUFFERS was very large.
	Also modified auto mode so that tests where
	reclen>filesize are skipped.

	Modified preadv code to vary the number of buffers as
	necessary such that they will fit in min(MAXBUFFERSIZE,filesize).
	This fixes problems where the number of buffers in
	the i/o vector exceeded the size of mainbuffer.

	Added bzero for buffer when it is first malloc'd.  This
	ensures that it is initialized before use.

	Created a script (profile.fs) that runs a series of tests
	to generate a "box" around common application variables
	such as filesize, buffer size, buffer encachement, and
	number of concurrent processes.  This is intended to serve
	as the "standard" filesystem profile.

	buffer reset to mainbuffer before each test loop

V2.3 (kcollins):

	added -F option to write to specify pathnames for throughput
	tests (allowing throughput tests to multiple filesystems).

V2.4 (capps):
	Changed preadv/pwritev to use a non-sequential access pattern.
	Changed the version number.
	Moved all user interface values to KB. This simplifies
	the user interface. (consistant scaling) and it also
	allows one to start with 512kb file. This is very important
	since the first indirect block causes a significant
	slowdown in the initial write cases.

V2.5 (capps):
	Re-structure and cleanup.

V2.6 (kcollins)
	Bug fix for the throughput tests.

V2.7 (capps):
	Added -o flag. This makes all file opens for writes
	have the O_SYNC flag set. This makes all writes go
	to disk before competion. This is useful for seeing
	what the media can do without the buffer cache helping.

V2.8 (capps):
	Added -V flag. This turns on pattern verification. If
	the user were to type:
	-V 165
	Then bit pattern 0xa5 would be placed in every byte in the
	buffer and when read back from buffer cache, or disk, 
	it will be verified to be correct. If it fails then
	the error handler will specify the byte location of the
	miscompare.

V2.9 (capps):
	Added fread/re-fread, fwrite/re-fwrite to list of tests.
	Added -E to allow the user to run pread and friends as an option.

V2.10 (capps):
	Added -R. This will generate Excel compatible files that
	can then be imported into Excel and graphed.
	Added support for 5 targets to the makefile.
	Added -M  This prints out the uname -a stuff about a machine.
	Added -O  This gives all results in operations/sec instead of KB/sec.
	More code cleanup. Update comments.

V2.11 (kcollins)
	added -A.  Auto mode with no crossover and read/write tests only
	changed default record size to 64KB (from 512 bytes)

V2.12 (capps)
	Added shared memory barrier sync for throughput mode. This
	provides much finer control over the actual timeing 
	of the children.
	Added mmap() for BSD (Convex) machines that do not 
	have System V shared memory.
	Added two ways of showing throughput results. The second
	method takes into consideration children that lag behind
	due to slow devices, and gives results that are more accurate.
	Cleanup of some tab problems in throughput results.
	Cleanup of floating point output taking to much space.
	Added -d to allow a variable delay comming out of the barrier 
	in the throughput tests.
V2.12	(kcollins)
	added declaration for create_list to make ansi c compiles work
	several fixes to some of the SPPUX 5.x make targets
	added date run to banner (hope this doesn't break your scripts $-)

V2.13	(capps)
	Added "stone walling". During throughput tests, if one process
	finishes then all others are sent a signal to tell them 
	to stop. (The parallel region has finished). This provides 
	better numbers for throughput.  
	Only bzero or fill min(reclen,CACHE_SIZE) this saves a bunch
	of paging on workstations with small memory systems.
	Fixed broken target in the makefile.
	Note: use of -d is not advised. It makes the children not run 
	in parallel.
V2.14   (capps)
	Bug fix to avoid anomaly in SPP-UX. In SPP-UX the filesystem
	code preallocates meta-data to improve initial file writes.
	The first indirect block allocation was causing a block 
	of zeros to be written syncronously. In SPP-UX the filesytem
	code preallocates zero filled blocks when the first writer
	touches a filesystem after a sync. A pool of on disk zero'd 
	blocks are created asynchronously and handed out to writers
	when they cross the boundry into the first level indirect
	and would have had to stop and wait for the zero filled 
	block to be written. Iozone's testing methodology was not
	allowing the OS to have any time to complete the async
	pre-allocation and was not showing the speed up that real
	applications would see.

V2.15   (capps)
	Improve throughput testing mode.

V2.16	(capps)
	Added -U option. This allows the filesystem to be unmounted
	and remounted between tests. This guarentees that the buffer
	cache is cold.

V2.17	(capps)
	Added -T option. This makes the throughput tests use 
	threads instead of processes.  Currently using pthread_create(),
	pthread_self(), and pthread_exit(). 
	Cleaned up file cleanup mechanism. Control C will now cause
	all temp files to be deleted.  Removed all signals used to
	control sub-processes.

V2.18	(capps)
	Cleanup. Added read stride, read backwards to the throughput
	tests. Various bug fixes

V2.19	(capps)
	Removed all calls to malloc() and all use of system V shared 
	memory. mmap() is much easier to deal with. As for malloc()
	HP programs are very limited on the ammount of malloc() space
	and not nearly so constrained on mmap() memory. It was necessary
	to move to mmap() since multiple threads all need buffers in
	the processes address space. 
	Removed dependency on first thread being number 2. Iozone now
	probes to find out what the thread library will return for 
	the first thread.  This makes the switching thread libraries
	much easier.

V2.20	(capps)
	Children now set stop_flag and shutdown all other children.There
	is no further need to tell the parent to distribute the stop_flag.
	verify, purge, and osync are now supported in the throughput
	tests. Fixed bug where pthreads stack size was causing 
	segmentation violation when purgeit() was called for buffer
	that were greater than 256kb.

V2.21	(capps)
	Enhanced throughput reporting. Now provides:
	Child throughput, Parent throughput, Minimum throughput for
	any child in the group, Maximum throughput for any child in the 
	group, and Minimum transfer count. Due to stone walling
	not all children write the full requested size. This 
	minimum transfer count provides the user with knowledge of 
	how much work was performed by the slowest child. 
	Added -C flag. This allows the user to see all of the transfer
	counts for each child. Had to add system 5 shared memory back.
	Linux does not support mmap(MAP_ANONYMOUS|MAP_SHARED). So it
	must use SYSV shared memory get get sharing working.

V2.22	(capps)
	Made changes to make iozone work correctly on Linux on a 
	PC. Changes are just scaling down the test to fit on a 
	pc, and scaling down shared segments to < 16 Meg so it
	can run on an Intel 386 class machine.
	Added: -L #   Set the processor cache line size in bytes.
	Added: -S #   Set the processor cache size in kbytes.
	Removed spin wait in parent waiting for threads to 
	finish each throughput test. Code not uses thread_join().
	Fixed -O (operations/sec) mode to work in throughput tests.

V2.23	(capps)
	Close small timing hole where thread/process has set stop flag
	and others are in a system call. The hole allowed threads/processes
	to continue to increment work done after one had finished and
	told the others to stop. The result was that the children would
	report slightly high numbers as they were not truely parallel
	at the finish line. Added random read throughput test.
	Fixes for VxFS small extents being created by prime_zb() functions.
	Provides more details about the throughput run.

V2.24	(capps)
	Added support for -R (Excell chart generation) to the throughput
	tests. Also added support for the -O (ops/sec) to the throughput
	Excell chart.

V2.25	(capps)
	Added support for selecting which test to run. -i #
	-i 0 -i 3 
	will run write and read-backwards tests only. For a list
	of the test numbers type iozone -h.
V2.26	(capps)
	Added support for LARGE_FILES for the hpux-11.0 target.
	

V2.27	(capps)
	All tests now verify one long word of data from each page
	written/read to/from the file. This is to level the 
	playing field with systems that do not move data 
	when "read" or "write" is called, but instead just
	map the file and perform the I/O when the address space
	is touched. Benchmarks that do not validate the data
	,at least touch each page, do not measure the read/write
	times just the map times.
	Note: The -V option still verifies each byte of the buffer,
	the default is now to verify one long from each page.

V2.28	(capps)
	Added support for benchmarking mmap() files.
	Added more command line options. -B -G -D
		B = Use mmap() files for the benchmark.
		G = Use msync(MS_SYNC) for mmap files.
		D = Use msync(MS_ASYNC) for mmap files.

V2.29	(capps)
	Bug fixes for:
		Combination of running individual tests and mmap() files support.
		Stride read bug that caused only portions of the total file to be 
		examined.
	
V2.30	(capps)
	Fixups for build under SPP-UX