Skip to contentSkip to conversation

huge read/write-operations are splitted by the kernel

blockdevices/NBD Sven Köhler 2 participants 6 messages Sep 8, 2003

lore
Share & tools
lore
#0 lore raw

Hi,

i discussed a problem of the NBD-protocl with Pavel Machek. The problem i saw is that there is no maximum for the length field in the requests that the NBD kernel module sends to the NBD server. Well, this length field is the length field from the read/write-operation that the kernel delegates to the blockdevice-implementation. I did some tests tests like
dd if=dev/nbd/0 of=/dev/null bs=10M
and our NBD-server implementation printed out the length field of each reqeust. There was a very regular pattern like
0x1fc00 (127KB)
0x00400 (1KB)
0x1fc00
0x00400
...
Well, can anybody explain that to me?
(why so "little" 1KB requests? but that's not important)

Well, i also tested
dd if=dev/nbd/0 of=/dev/null bs=1
which means that the device will be read in chunks of 1byte. The result was the same: 127KB, 1KB, 127KB, 1KB...

I guess the caching layer is inbetween, and will devide the huge 10MB requests into smaller 127KB ones, as well as joining the small 1byte requests by using read-ahead i guess. Perhaps you could tell me how i can turn off caching. Than i will test again without the cache.

The thing i want to know is, if there is any part of the kernel that gaarantees that a read/write requests will not be bigger that a certain value. If there is no such upper limit, the NBD itself would need to split things up which might become a complicated task. This task need to be done, because it can become very difficult for the NBD server to handle huge values, and one huge requests will block all other pending small ones due to limitations of the NBD protocol.

Thx
Sven

#1 lore raw
Quoted 30 lines · Sven Köhler

Hi,

i discussed a problem of the NBD-protocl with Pavel Machek. The problem i saw is that there is no maximum for the length field in the requests that the NBD kernel module sends to the NBD server. Well, this length field is the length field from the read/write-operation that the kernel delegates to the blockdevice-implementation.
I did some tests tests like
dd if=dev/nbd/0 of=/dev/null bs=10M
and our NBD-server implementation printed out the length field of each reqeust. There was a very regular pattern like
0x1fc00 (127KB)
0x00400 (1KB)
0x1fc00
0x00400
...
Well, can anybody explain that to me?
(why so "little" 1KB requests? but that's not important)

Well, i also tested
dd if=dev/nbd/0 of=/dev/null bs=1
which means that the device will be read in chunks of 1byte. The result was the same: 127KB, 1KB, 127KB, 1KB...

I guess the caching layer is inbetween, and will devide the huge 10MB requests into smaller 127KB ones, as well as joining the small 1byte requests by using read-ahead i guess.
Perhaps you could tell me how i can turn off caching. Than i will test again without the cache.

Quoted 7 lines

The thing i want to know is, if there is any part of the kernel that gaarantees that a read/write requests will not be bigger that a certain value. If there is no such upper limit, the NBD itself would need to split things up which might become a complicated task. This task need to be done, because it can become very difficult for the NBD server to handle huge values, and one huge requests will block all other pending small ones due to limitations of the NBD protocol.

You'll probably find that if you bump the max_sectors count if your drive to 256 from 255 (that is the default if you haven't set it), then you'll see 128kb chunks all the time.

See max_sectors[] array.

Signature · 1 line
Jens Axboe
#2 lore raw
Quoted 3 lines

You'll probably find that if you bump the max_sectors count if your drive to 256 from 255 (that is the default if you haven't set it), then you'll see 128kb chunks all the time.

Why is 255 the default. It seems to be an inefficient value. Perhaps the NBD itself should set it to 256.

Quoted 1 line

See max_sectors[] array.

Well, i found the declaration, but i can't imagine how to set the values in it.

#3 lore raw
Quoted 6 lines · Sven Köhler
Quoted 3 lines

You'll probably find that if you bump the max_sectors count if your drive to 256 from 255 (that is the default if you haven't set it), then you'll see 128kb chunks all the time.

Why is 255 the default. It seems to be an inefficient value. Perhaps the NBD itself should set it to 256.

To avoid 8-bit wrap arounds, basically. Not sure it's still very valid, you are free to compile your kernel with it set to 256. 2.6 uses 256 by default.

Quoted 4 lines
Quoted 1 line

See max_sectors[] array.

Well, i found the declaration, but i can't imagine how to set the values in it.

You can grep for other examples in the kernel, I would imagine?

Signature · 1 line
Jens Axboe
#4 lore raw
In reply to Jens Axboe
Quoted 5 lines

You'll probably find that if you bump the max_sectors count if your drive to 256 from 255 (that is the default if you haven't set it), then you'll see 128kb chunks all the time.

See max_sectors[] array.

To make it clear:
the kernel will never read or write more sectors at once than specified in the max_sectors array (where every device has its own value), right?

#5 lore raw
Quoted 9 lines · Sven Köhler
Quoted 5 lines

You'll probably find that if you bump the max_sectors count if your drive to 256 from 255 (that is the default if you haven't set it), then you'll see 128kb chunks all the time.

See max_sectors[] array.

To make it clear:
the kernel will never read or write more sectors at once than specified in the max_sectors array (where every device has its own value), right?

Correct

Signature · 1 line
Jens Axboe