Hi,
Does mount point specification while partitioning (order in which I specify /, /boot, swap etc..) affect performance? I am not sure about the syntax, but I guess one can also specify address/block range while partitioning. Does it affect IO performance? Probably a stupid question, but just curious.. Any insights?
Thanks, CS.
Carlos Santana wrote:
Hi,
Does mount point specification while partitioning (order in which I specify /, /boot, swap etc..) affect performance? I am not sure about the syntax, but I guess one can also specify address/block range while partitioning. Does it affect IO performance? Probably a stupid question, but just curious.. Any insights?
Not a stupid question at all. For ordinary disk drives the answer is yes, absolutely. The outer tracks of a disk are physically longer, and any but the most ancient of disk drives will pack more sectors into those tracks. Since the disk rotates at a constant RPM, more sectors per second pass under the head on the outer tracks. The ratio of data rates for the outermost vs. innermost tracks is typically 2:1 or a bit higher. Add to this the need for more and longer seeks for filesystems on the inner tracks (again, less data on each physical track), and the performance degrades even more.
On most disks cylinder numbering starts at the outer tracks, but I have heard of disks that number their cylinders in the opposite direction -- never actually seen one, though.
On Dec 22, 2009, at 6:13 PM, Robert Nichols rnicholsNOSPAM@comcast.net wrote:
Carlos Santana wrote:
Hi,
Does mount point specification while partitioning (order in which I specify /, /boot, swap etc..) affect performance? I am not sure about the syntax, but I guess one can also specify address/block range while partitioning. Does it affect IO performance? Probably a stupid question, but just curious.. Any insights?
Not a stupid question at all. For ordinary disk drives the answer is yes, absolutely. The outer tracks of a disk are physically longer, and any but the most ancient of disk drives will pack more sectors into those tracks. Since the disk rotates at a constant RPM, more sectors per second pass under the head on the outer tracks. The ratio of data rates for the outermost vs. innermost tracks is typically 2:1 or a bit higher. Add to this the need for more and longer seeks for filesystems on the inner tracks (again, less data on each physical track), and the performance degrades even more.
On most disks cylinder numbering starts at the outer tracks, but I have heard of disks that number their cylinders in the opposite direction -- never actually seen one, though.
Also, for random IO the opposite is true, the rotational latency is significantly smaller on the inner tracks than the outer tracks, so random OPs perform better there.
Though having different workloads on opposite sides of the disk is counter productive, but say you had one large volume for random IO workloads and another large volume for sequential workloads, you could allocate the beginning chunk to your most performance oriented servers on the sequential volume and the end chunk to your most performance oriented servers on the random volume.
-Ross
Ross Walker wrote:
Also, for random IO the opposite is true, the rotational latency is significantly smaller on the inner tracks than the outer tracks, so random OPs perform better there.
um, most all hard disks are CAV, so the rotational latency measured in milliseconds is constant throughout the disk. usually 50% of a turn is the assumed mean rotational latency
On Dec 22, 2009, at 10:14 PM, John R Pierce pierce@hogranch.com wrote:
Ross Walker wrote:
Also, for random IO the opposite is true, the rotational latency is significantly smaller on the inner tracks than the outer tracks, so random OPs perform better there.
um, most all hard disks are CAV, so the rotational latency measured in milliseconds is constant throughout the disk. usually 50% of a turn is the assumed mean rotational latency
I think you might be confusing CAV with CLV of optical drives.
http://en.m.wikipedia.org/wiki/Constant_Angular_Velocity
-Ross
On Wed, 2009-12-23 at 09:37 -0500, Ross Walker wrote:
<snip>
I think you might be confusing CAV with CLV of optical drives.
http://en.m.wikipedia.org/wiki/Constant_Angular_Velocity
-Ross
That was my thought. However, I think most are missing the boat on this.
I have always looked at the anticipated work profile on the drive and tried to place partitions to minimize seek time - that being the single biggest latency issue, IMO. By placing the most frequently accessed partitons adjacent to each other, and near the middle of the platter(s), seek delays are minimized.
With the advent of LVM, I feel this is more easily fine tuned, initially and later after the /real/ workload can be statistically profiled.
Ross Walker wrote:
I think you might be confusing CAV with CLV of optical drives. http://en.m.wikipedia.org/wiki/Constant_Angular_Velocity
no, I'm not. most HD's ('green drives' complicate this some) spin at a constant RPM, so the rotational latency is the same on the inner and outer tracks, an average of 1/2 turn, about 4mS for a 7200 rpm drive, and 2mS for a 15000rpm enterprise drive . However, the data rate changes. so the outer tracks have more data on them, which is read at a higher speed in megabytes/second
On 12/23/2009 07:29 PM, John R Pierce wrote:
Ross Walker wrote:
I think you might be confusing CAV with CLV of optical drives. http://en.m.wikipedia.org/wiki/Constant_Angular_Velocity
no, I'm not. most HD's ('green drives' complicate this some) spin at a constant RPM, so the rotational latency is the same on the inner and outer tracks, an average of 1/2 turn, about 4mS for a 7200 rpm drive, and 2mS for a 15000rpm enterprise drive . However, the data rate changes. so the outer tracks have more data on them, which is read at a higher speed in megabytes/second
That's why in ancient times one was setting up partitions so that the swap area was the the beginning (mostly the outer tracks of the HD -- never hit a drive that did it the other way round) of the drive.
Try it yourself, get a spare HD and create three partitions on it, two smaller ones at beginning/end of the drive, the third one filling the gap between them; install bonnie++ and compare the transfer rates.
Timo
Timo Schoeler wrote:
On 12/23/2009 07:29 PM, John R Pierce wrote:
Ross Walker wrote:
I think you might be confusing CAV with CLV of optical drives. http://en.m.wikipedia.org/wiki/Constant_Angular_Velocity
no, I'm not. most HD's ('green drives' complicate this some) spin at a constant RPM, so the rotational latency is the same on the inner and outer tracks, an average of 1/2 turn, about 4mS for a 7200 rpm drive, and 2mS for a 15000rpm enterprise drive . However, the data rate changes. so the outer tracks have more data on them, which is read at a higher speed in megabytes/second
That's why in ancient times one was setting up partitions so that the swap area was the the beginning (mostly the outer tracks of the HD -- never hit a drive that did it the other way round) of the drive.
Try it yourself, get a spare HD and create three partitions on it, two smaller ones at beginning/end of the drive, the third one filling the gap between them; install bonnie++ and compare the transfer rates.
But these days, nothing should ever be reading from swap, although you might write a bit there. If it does, buy some more RAM instead of worrying about disk performance.
On 12/23/2009 08:15 PM, Les Mikesell wrote:
Timo Schoeler wrote:
On 12/23/2009 07:29 PM, John R Pierce wrote:
Ross Walker wrote:
I think you might be confusing CAV with CLV of optical drives. http://en.m.wikipedia.org/wiki/Constant_Angular_Velocity
no, I'm not. most HD's ('green drives' complicate this some) spin at a constant RPM, so the rotational latency is the same on the inner and outer tracks, an average of 1/2 turn, about 4mS for a 7200 rpm drive, and 2mS for a 15000rpm enterprise drive . However, the data rate changes. so the outer tracks have more data on them, which is read at a higher speed in megabytes/second
That's why in ancient times one was setting up partitions so that the swap area was the the beginning (mostly the outer tracks of the HD -- never hit a drive that did it the other way round) of the drive.
Try it yourself, get a spare HD and create three partitions on it, two smaller ones at beginning/end of the drive, the third one filling the gap between them; install bonnie++ and compare the transfer rates.
But these days, nothing should ever be reading from swap, although you might write a bit there. If it does, buy some more RAM instead of worrying about disk performance.
Sure, absolutely no question; *but* in the (ancient) times it was important, it was 'nice' to have it as fast as possible, i.e. on the fastest section(s) of the used HDs. So...
Timo
Timo Schoeler wrote:
But these days, nothing should ever be reading from swap, although you might write a bit there. If it does, buy some more RAM instead of worrying about disk performance.
Sure, absolutely no question; *but* in the (ancient) times it was important, it was 'nice' to have it as fast as possible, i.e. on the fastest section(s) of the used HDs. So...
unless you're using tmpfs, where your swap space doubles as backing store for your /tmp I use this quite regularly on solaris, it performs much better than a conventional journaled file system, as recovery from system crashes is totally not a priority.
On Dec 23, 2009, at 1:29 PM, John R Pierce pierce@hogranch.com wrote:
Ross Walker wrote:
I think you might be confusing CAV with CLV of optical drives. http://en.m.wikipedia.org/wiki/Constant_Angular_Velocity
no, I'm not. most HD's ('green drives' complicate this some) spin at a constant RPM, so the rotational latency is the same on the inner and outer tracks, an average of 1/2 turn, about 4mS for a 7200 rpm drive, and 2mS for a 15000rpm enterprise drive . However, the data rate changes. so the outer tracks have more data on them, which is read at a higher speed in megabytes/second
You know your right. I don't know what I was thinking, a rotation is a rotation and if it takes 4ms on the outer tracks then it takes 4ms on the inner tracks.
It is I who had to two mixed up. It would be CLV that would make the disk spin faster as it approached the inner tracks which is the only way rotational latency would decrease.
Sorry for the noise, we now continue with your regularly scheduled program.
-Ross
-
Carlos Santana -
John R Pierce -
Les Mikesell -
Robert Nichols -
Ross Walker -
Timo Schoeler -
William L. Maltby