Ask the Expert: Changing The Database Rules For Performance With XtremIO

Roberto,

Thanks for posting this the event! We really look forward to discussing Oracle on XtremIO. We are also very excited to have Matt Kaberlein from the Oracle Specialist team leading the event as he has a wealth of experience in working closely with customers on using EMC's XtremIO all-flash array.

Very much looking forward to supporting this conversation, here's another tweet with an image attached:

Ask the Expert: Changing #Oracle Database performance rules w/ #EMC XtremIO http://bit.ly/1BNjpba pic.twitter.com/QYbhueT1aJ

So, I’m going to get the discussion started by talking about the importance of array testing. All flash arrays like XtremIO demand different approaches to performance testing because of features like inline data reduction, all zero optimization, and compression. Tools like btest and vdbench enables customers to run performance tests with unique data so array features like inline data reduction and zero optimization don’t impact the test results. My question is: How important is it to customers that performance tests show the most strenuous performance scenario?

Here is a short independent blog by Crispyire called Everything is so... Xtrem that touches upon the same subject of testing.

Welcome everyone, this ATE discussion has began. Feel free to post your questions and our SMEs will be sure to answer them before the end of our event. Let's keep this discussion respectful and informative. Thanks!

Sam:

I would have to say the zero performance impact snapshots.

Regards,

Jeff

George:

I have seen that argument, but with the ability to create a virtually unlimited number of zero performance cost copies, with very little capacity overhead (at least if you delete them fairly quickly), makes XtremIO irresistible for making a test/dev farm.

Regards,

Jeff  

What is your favorite XtremIO feature?

Oracle Heretic,

My vote would be the inline compression. As a DBA you get immediate space savings somewhere in the range of 1.5 to 2 times. An Oracle database 2TB in size could compress down to 1TB. This means the DBA can fit more databases onto XtremIO.

The other part of the compression is with snapsots. Over time most snapshots have a data delta from the source which means they take up space. So compression offers an advantage here too because the data delta is compressed. This means more snapshots are possible.

Finally, this could save the customer money in that they don't have to purchase Advanced Compression.

George,

Good question. In talking to a few DBAs the question they asked, "Should XtremIO be reserved for production databases?" Nothing replaces proper sizing to verify but generally XtremIO is great for non-production (test, development, and sandbox) databases. Here is how I think about it:

• Virtually unlimited number of zero performance cost snapshots. Yes, I’m borrowing from Oracle Heretic because he has it right. This means the storage administrator can instantaneously create snapshots of Oracle database to give to the DBA. These snapshots initially take no space and have no performance impact on production.
• Compression is my favorite XtremIO feature because snapshots benefit too. For example, let’s say a data delta of 20% across all your snapshots equals 25TB of uncompressed space. That’s a lot of space dedicated to snapshots. Assuming a compression of 2X the snapshots now take 12.5TB which is a big capacity savings!

The speed and simplicity of using XtremIO snapshots together with compression make this storage platform ideal for non-production databases.

Matt Kaberlein in this video does a good job talking about XtremIO snapshots: fast forward to the 19 minute 50 second mark.

George

How does XtremIO benefit non-production workloads – is it even worth it for them?  Or should we only use for production?

Per your question, a couple of months back I was working with a group of Oracle & SQL Server DBAs at a large Oracle & SQL Server customer to understand their I/T challenges and then map these to XtremIO capabilities.

Obviously, the DBAs liked the idea of fast consistent response time for small & bigger block read & write I/O for their Prod DBs. ie. DBWR & LGWR I/Os

But, one of their Apps needs 10-14 clones of a 2TB+ DB in Dev/Test for their end users for a variety of reasons. ie training, Developer software changes, Functional testing, a good ole fashion DBA instance to test patches & such, ... So, being able to leverage dedupe & compression to reduce the storage foot print was very appealing to them. And, since they can leverage XtremIO Snapshots or RMAN Duplicate to create the Clones & get the dedupe/compression this was icing on the cake for them.

Recently, Oracle has been posting some competitive metrics comparing XtremIO to the FS1. I’m going to focus on Oracle’s test method. In most of competitive slide decks there is a slide showing XtremIO performance metrics in a blue bar graph and Oracle’s metrics in a red bar graph. At the bottom of the slide in 8 point font, perhaps even smaller, is the statement “Oracle performance internal testing.” So there is no way for customers for read how the testing was conducted. Recently, Oracle did have a webcast and provided more support for their comparison metrics. In one of the slides they had a list titled, “A few points on this chart” that stated “The test is simple IOMETER with appropriate parameters of BOTH FS1 and EMC. No hocus pocus, no BS.” This is FALSE!

At EMC we use btest not IOMETER. I recommend reading, “Considerations for Testing All-Flash Array Performance: What you need to know when evaluating the performance of Solid State Storage Systems.” At the bottom of page 10 we talk about IOmeter…

“The most commonly used option. It’s attractive because it has a GUI, which makes it simple to get started. But it’s also easy to use IOmeter to generate inaccurate results, and building complex testing scripts can be more challenging.”

Just my opinion but IOmeter is not what you want to use for internal testing of an all-flash array. What does the same white paper have to say about BTEST?

“High performance and low overhead load generator for Linux. It doesn’t include all the options and built-in automation capabilities of some of the other tools. It can generate a very high multi-threaded load with a specific deduplication ration. BTEST is a command line tool that’s great for running quick manual tests or as the foundation of custom automation”

Reads like a better tool and the part about being able to tune the test for a deduplication ratio is very important. Because XtremIO is an all-flash array (can’t be configured any other way) it’s important to correct for inaccurate results. For example, inline data reduction, all zero optimization and compression if utilized while testing can generate inaccurate results. So in terms of performance testing all-flash arrays tools like btest and VDbench are preferred.

There is a great third party paper by Principled Technologies called, “Scaling Oracle 12c Database Performance with EMC XtremIO Storage in a Database as a Service Environment” that shows some real test results with all the steps used to generate the results.

Hi Sam, I would prefer XtremIO's Data Protection (XDP).  It combines the most advantages of traditional RAID algorithms and avoid their pitfalls.  Past RAID algorithms had to consider how to keep data contiguous, so as to avoid disk drive head seeks.  The XDP algorithm places and accesses data in any location on any SSD.  XDP presumes random access media, such as Flash, is present in the array, and it can lay out data and read it back in highly efficient ways which would heavily penalize in a traditional disk-based RAID algorithm, but with no negative effects in XtremIO all-Flash architecture.

XDP significantly enhances the endurance of the underlying Flash media compared to any previous RAID algorithm.  This is an important consideration for cost saving.

Each XtremIO X-Brick contains one DAE with 25 SSDs. The SSDs are configured with N+2 row and diagonal parity, as shown here. If one of the SSDs in an X-Brick fails, XDP automatically rebuilds the data to unused space on the remaining SSDs, using distributed parity and data on the remaining SSDs, while dynamically reconfiguring incoming new writes into a 22+2 stripe size to maintain N+2 double failure protection for all new data written to the array. XDP always maintains a reserved capacity on each drive to perform this drive rebuild.

Zhao2,

I agree data protection is very important but particularly for Oracle databases. DBA are rightfully concerned about availability, recoverability, and capacity overhead.

XtremIO Availability:

From the hardware perspective, no component is a single point of failure as the storage controller, DAE and InifiniBand switch in the system is equipped with dual power supplies. The software architecture is built in a similar way. Every piece of information that is not committed to the SSD is kept in multiple locations, called Journals. To read more about XtremIO High Availability and many other features we recommend, Introduction to the EMC XtremIO Storage Array 3.0.

Recoverability:

There is a very good XtremIO paper called, "XtremIO Data Protection (XDP) Flash-Specific Data Protection, Provided by XtremIO" that takes the reader through a detailed review of XDP. XtremIO provides a superior rebuild process. When a traditional RAID 6 array faces a single disk failure, it uses RAID 5 methods to rebuild it by reading each stripe and computing the missing cell from the other cells in the stripe. In contrast, XtremIO uses both the P and Q parity to rebuild the missing information and uses an elaborated algorithm that reads only the needed information for the next cell rebuild.

Below is a comparison of XDP reads for rebuilding a failed disk with those of different RAID Schemes (from the above paper).

XDP's efficiency advantage leads both to faster rebuild times as well as better array performance during rebuilds since fewer back-end I/Os are required to complete the rebuild, leaving more front-end I/Os available for user data.

Capacity Overhead:

Be careful of the statement, “Engineered for Flash” because if the array is using traditional RAID it is using more capacity overhead for flash! In this video Ask The Expert: Changing The Database Rules For Performance With XtremIO at the 24 minute 53 second mark Matt Kaberlein talks about XDP. The table from the video shows the differences in capacity overhead:

DBAs will be please to know their investment in XtremIO gives them the most flash capacity because it has the least overhead (8%) compared to RAID 1, 5, or 6. In terms of availability, recoverability, and reduction in capacity overhead XtermIO is a flash leader and others arrays using tradition RAID, haven’t been engineered for flash, at least at the data protection level.

Hey George,

Should you use ASM redundancy on all-flash?

What you would do is use ASM external redundancy to enable the XtremIO array to be responsible for data protection (ie.XDP). 

Also, with using ASM, use the default ASM file template's AU and AU stripe size. No need to change with these.