Name Game

[Angus Crome]

In reference to the below posts (and others);

http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0xdf0a31ec5e34d711abdc0090277a778c,00.html

I propose we redefine "using varied PV Links", as load balancing.
Anyone out there want to come up with a new name?

I propose we start referring to it as "load sharing" (since that is what it really is). Load balancing implies some operative action be done to keep IO/data throughput equal.

I'll give 1 point for any valid answer. I know it is kind of silly, but my CS background just won't let it be.

[RAC_1]

Alternate links are for redundancy and not for load balancing. If one fails the alternate path taker over. That means data going or coming from disk is from one path only.

This is done with vgextend and vgreduce commands.

For load balancing - i.e to flow data through both paths you will have to use Secure path. Secure path takes care of load balancing + redundancy and system sees just one device file. So confusion with two device files is avoided.

[Sridhar Bhaskarla]

I won't mind if you call it anything as long as one understands it.

Anil - Take the following example.

one lv is lvmstriped across four disks c1t0d0, c1t0d1, c1t0d2, c1t0d3 with a stripe size of 64k.

Now I configure the vg in such a way that the order in /etc/lvmtab looks like this.

c1t0d0
c2t0d1
c1t0d2
c2t0d3
c2t0d0
c1t0d1
c2t0d2
c1t0d3

This does load balance (or share to make you comfortable) the IO. This provides redundancy as well.

Your statement implies that they are not to be used for loadbalancing which is incorrect. Even in the above case, the other path will take over with the auto switch feature of LVM.

However, every logical volume is not to be striped. In which case, the load sharing may not be effective. But you can still alter the paths and make use of both the paths to the best you can.

-Sri

[RAC_1]

For Shridhar,

Thanks. One query though.

Load balancing will happen if I strip a lvol onto diff disks. right?

Say I have two controlers-c1 and c2

I create lvol stripped onto two disks. My lvmtab looks like this
vg01
/dev/dsk/c1t1d0
/dev/dsk/c1t1d1
/dev/dsk/c2t1d0
/dev/dsk/c2t1t1

What will happen in this case?

Not clear on this-particulary stripping and how load balancing is done in this case?

As I always say - there is something new to learn.

[Sridhar Bhaskarla]

Anil,

In this particular case, there will be no load sharing. It is the order in /etc/lvmtab that matters. (that you see in vgdisplay -v). Here both c1's are defined as primaries so all the IOs will go through that controller.

With a little tweaking of your configuration, you can make use of alternate link online. Do this way.

vgreduce vg01 /dev/dsk/c1t1d1
vgextend vg01 /dev/dsk/c1t1d1

This will change your lvmtab to look like

vg01
/dev/dsk/c1t1d0
/dev/dsk/c2t1d1
/dev/dsk/c2t1d0
/dev/dsk/c1t1t1

This will try to balance the load across both the controllers. If you do sar -d 2 20, you can observe IOs going through both paths.

-Sri

PS: Even if you get 90:10 share, it is better than 100:0. It does matter in companies like mine where every millisecond response directly translates into dollars.

[RAC_1]

Thanks shridhar

[Angus Crome]

OK, Sridhar and Anil, this wasn't really the point, but I will make one comment on the discussion (I'll also give you a total of 1 point each).

Anil, I believe you are reading the point of varied PV links as trying to balance the IO's across a primary and alternate for the same disk. You are correct in saying this is not possible with PV Links.
What you do instead (and the point of this thread) is share the combined IO to all disks, somewhat equally, over how ever many channels for which you have primary and alternate disks. You just adjust which channel each primary actually sits on. Some amount of IO traverses each channel this way, but it has no guarantee about how much (certainly not equal) at any given time.

[Sridhar Bhaskarla]

Angus,

You probably cleared out what others have been saying about "not for load balancing". But even in a true load balance scenario as theorized by you, unless the disk is fast enough to yield throughput more than the pipes can handle, that way of load balancing may not help much.

I was only trying to be careful about the fact that the user community may derive from these comments that "they should not use alternate links for load sharing".

Now this thread may get more messages with expert opinions on what these links are for etc., etc.,. But I care less about the theories as far I can make use of my resources to the best I can.

-Sri

[Angus Crome]

Absolutely Sridhar, I totally agree. It does not make one iota of difference what we call it, as long as it can be utilized to increase performance (even 1%). Any gain is a good gain, when it is so easy to do.

[Angus Crome]

Guess no one is interested.

Consider this thread closed....

[Robert-Jan Goossens]

think you are intrested, but dont know if the mail account is still active.

http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=147294

Good weekend,
Robert-Jan
0 Point please

[Todd McDaniel_1]

lets call PVlinks what it is...

PVLINKS=failover path. As was stated well above... the ALT path lays there like a dead mackerel and doesnt do a darn thing till the Primary path dies or is removed.

And that is it...


The only way I know of that you can get anykind of psuedo-loadbalancing and it is STILL not loadbalancing, is to alternate your primary and alternate paths between your controllers within a VG or LVOL.

This is merely dividing the work across multiple controllers which might give you a slight increase in perforance.

The only way to get true loadbalancing is to use a 3rd party product like Securpath or Powerpath, so far as I know.