This post is more than 5 years old
53 Posts
0
1643
How many meta sizes do you have in a VMAX thin environment?
All:
I've been told that the optimal number of meta members in a striped meta comprises multiples of 8. This is supposed to give best concurrency because of the way the VMAX works.
No problem so far. But I just had a customer request a LUN that would only comprise 2 of my metas at my current auto-meta size.
My installing engineer said I should minimize the number of different meta sizes throughout the box. My assumption is b/c it becomes unwieldy and hard to keep track of a ton of different sizes.
Having said that, what are you guys all doing? Do you even concern yourself with making 8 way metas for performance reasons? If so, have you tested an 8 way against a 2 way to see what kind of difference you notice? I'm curious if it's significant.
Also, are you creating custom metas for every request to match the customers' sizing requirements?
Thanks!!!!
Brandon
dynamox
2 Intern
2 Intern
•
20.4K Posts
0
March 1st, 2012 09:00
our "building block" device size is 17263 MB, we use these devices to create striped Meta. We have so many customers/applications that it's very hard to use the same rule, we just try to use an even number.
BRiley1
53 Posts
0
March 1st, 2012 09:00
So after posting, I ran some tests. A 2 way meta shows 11% less IOPS on a SQLish IOmeter test than an 8 way of the same size and pool.
Just thought anyone reading this thread might want to know those results.
BRiley1
53 Posts
0
March 1st, 2012 10:00
Kiranghag:
This must be the EMC company line. I have been told the same. My test results completely refute this.
I'm sure EMC will say "it depends", and really it does. But on a pure performance driven write heavy workload with multiple threads, the concurrency I get from striped helps quite a bit. I doubt most people would ever run up against this limit, but there is a per device WP limit, which means you can run out of cache faster writing to a concurrent meta. Read cache should be the same regardless.
HTH,
Brandon
kiranghag
8 Posts
0
March 1st, 2012 10:00
may be little out of topic, but i have been told that striped metas are no longer necessary for higher performance in vmax thin pools. concatenation provides necessary performance.
have you seen this behaviour?
codyhosterman
286 Posts
1
March 1st, 2012 10:00
Symmetrix engineering actually recommends Striped in most cases for Virtual Provisioning for a couple reasons one which Brandon mentioned. The technical note for Symmetrix VP notes this and can be found on Powerlink:
http://powerlink.emc.com/km/live1/en_US/Offering_Technical/Technical_Documentation/300-006-718.pdhttp://powerlink.emc.com/km/live1/en_US/Offering_Technical/Technical_Documentation/300-006-718.pdf
Here is the relevant excerpt though:
In most cases, EMC recommends using striped rather than concatenated metadevices with Virtual Provisioning because they offer the following advantages:
RobertDudley
448 Posts
0
March 1st, 2012 11:00
The "it depends" answer is usually right. A striped meta will provide better performance than a concatenated however a concatenated is easier to expand. Its also application dependent a striped meta can perform better but it doesn't mean that you will always see that.
I usually leave it at if performance is of greater concern stripe it, if being able to expand is a concern concatenate it.
BRiley1
53 Posts
0
March 1st, 2012 15:00
The 2 member meta was the same size as the 8 member. I didn't test a single member, but I would think it would have worse performance at the limits than the 2. Again, these are IOMeter tests, and not really representative of a true workload.
vforde
67 Posts
0
March 1st, 2012 15:00
So does it follow through that one large Tdev equal in size to the meta (8/2 member), both from same pool would have a worse performance than the thin meta ? (assuming dev size is not greater than 240GB)
Quincy561
1.3K Posts
0
March 1st, 2012 17:00
A single TDEV device will have lower performance than a striped TDEV device. The performance will improve as you add members with a plateau around 4 members, with 8 probably giving slightly more due to the standard distribution of IO. Even with 100% cache operations (reads or writes).
AndyWOhio
21 Posts
1
March 2nd, 2012 14:00
When we implemented our first VMax about 2 years ago, the recommendation from EMC for meta devices was to just use concatenation for TDEVs since the thin pool itself is striped.
Since that time, EMC has discovered that striped metas can have performance advantages for high IOs because the IOs get to take advantage of the queue depths of all the devices. In the EMC VMax performance class I took recently, the class instructor was still touting the old policy, but an EMC performance engineer in the class mentioned that he standard has changed.