Hello.
I have an application that writes chunks of data to a file by using
pwrite. I have multiple threads, where one of the threads is writing
out 16k chunks of data, and the other thread is writing out 4k chunks.
I have been using dtrace and iosnoop to look at the application, and I
find it strange that my 16k write has been split up to two IO
operations. (I have added entry/return-probes for pwrite here)
UID PID D BLOCK SIZE COMM PATHNAME
pwrite64(8, 0x975d000, 16384, 0x3fff000)
pwrite64(9, 0xfc401000, 4096, 0x3000)
500 10909 W 3867928 12288 chunktest <none>
500 10909 W 3867952 4096 chunktest <none>
pwrite64 = 16384
500 10909 W 138184 4096 chunktest <none>
pwrite64 = 4096
I don't know much about a filesystem and block layout, but I thought
that if the blocks I was about to write are contigenous it may be
performed as a single IO operation? (3867928+(12288/512) == 3867952)
I have tried to use iosnoop to look on mkfile:
[...]
write(3, 0x8062730, 8192, 0x0)
write(3, 0x8062730, 8192, 0x0)
write(3, 0x8062730, 8192, 0x0)
500 25769 W 2223232 1048576 mkfile /export/home/mkfile
Now all of the small writes has been joined into one big IO operation.
Could anybody please enlighten me? Is there anything I can do to "optimize" my
IO to avoid the splitting of my 16k write into one 12k and one 4k?
Trond
This message posted from opensolaris.org
_______________________________________________
ufs-discuss mailing list
[email protected]
