atom feed21 messages in org.freebsd.freebsd-fsRe: Very inconsistent (read) speed on...
FromSent OnAttachments
Lev SerebryakovAug 30, 2011 12:10 pm 
Lev SerebryakovAug 30, 2011 12:17 pm 
Kirk McKusickAug 30, 2011 1:08 pm 
Lev SerebryakovAug 30, 2011 3:13 pm 
Lev SerebryakovAug 30, 2011 3:29 pm 
Lev SerebryakovAug 30, 2011 3:31 pm 
Bob FriesenhahnAug 30, 2011 3:39 pm 
Kirk McKusickAug 30, 2011 4:00 pm 
Jeremy ChadwickAug 30, 2011 5:42 pm 
Daniel KalchevAug 31, 2011 12:10 am 
Lev SerebryakovAug 31, 2011 12:38 am 
Lev SerebryakovAug 31, 2011 1:02 am 
Lev SerebryakovAug 31, 2011 1:10 am 
Lev SerebryakovAug 31, 2011 1:18 am 
Lev SerebryakovAug 31, 2011 1:36 am 
Daniel KalchevAug 31, 2011 1:48 am 
Lev SerebryakovAug 31, 2011 2:03 am 
Jeremy ChadwickAug 31, 2011 3:11 am 
Lev SerebryakovAug 31, 2011 4:36 am 
Daniel KalchevAug 31, 2011 4:46 am 
Lev SerebryakovAug 31, 2011 5:49 am 
Subject:Re: Very inconsistent (read) speed on UFS2
From:Lev Serebryakov (le@serebryakov.spb.ru)
Date:Aug 31, 2011 12:38:07 am
List:org.freebsd.freebsd-fs

Hello, Bob. You wrote 31 августа 2011 г., 2:40:07:

I'll try this experiment with mmap() and touching every 4096-th byte of mapped memory instead of read(2).

Strange enough, it gives only 40-50MiB/s and results are very consistent.

It really surprise me. I didn't think, that there will be so much difference, I was sure, that it will be almost equivalent speed.

FreeBSD does not seem to default to sequential read-ahead when memory mapping is used with sequential page access. Try using madvise() with the MADV_SEQUENTIAL option and see if it helps.

It were results with MADV_SEQUENTIAL. Code looks like this: (error checking is skipped here, but not in real code, of course):

fd = open(fileName, O_RDONLY | O_DIRECT); buf = mmap(NULL, fileSize, PROT_READ, 0, fd, 0); madvise(buf, fileSize, MADV_SEQUENTIAL); gettimeofday(&start, NULL); for (rd = 0; rd < fileSize; rd += 4096) c = buf[rd]; gettimeofday(&end, NULL); munmap(buf, fileSize); close(fd);

There are also MADV_WILLNEED, MADV_DONTNEED, and MADV_FREE. Careful use of these options can help performance quite a lot when data is large compared to memory.

It is too complex for simple linear read test :)

-- // Black Lion AKA Lev Serebryakov <le@serebryakov.spb.ru>

_______________________________________________ free@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-fs To unsubscribe, send any mail to "free@freebsd.org"