Kevin Closson

Subscribe to Kevin Closson feed Kevin Closson
Platform, Database and Storage Topics
Updated: 7 hours 26 min ago

You Scratch Your Head And Ponder Why It Is You Go With Maximum Core Count Xeons. I Can’t Explain That, But This Might Help.

Tue, 2016-06-14 00:36

Folks that have read my blog for very long know that I routinely point out that Intel Xeon processors with fewer cores (albeit same TDP) get more throughput per core. Recently I had the opportunity to do some testing of a 2-socket host with 6-core Haswell EP Xeons (E5-2643v3) connected to networked all-flash storage. This post is about host capability so I won’t be elaborating on the storage. I’ll say that it was block storage, all-flash and networked.

Even though I test myriads of systems with modern Xeons it isn’t often I get to test the top-bin parts that aren’t core-packed.  The Haswell EP line offers up to 18-core parts in a 145w CPU.  This 6-core part is 135w and all cores clock up to 3.7GHz–not that clock speed is absolutely critical for Oracle Database performance mind you.

Taking It For a Spin

When testing for Oracle OLTP performance the first thing to do is measure the platform’s ability to deliver random single-block reads (db file sequential read). To do so I loaded 1TB scale SLOB 2.3 in the single-schema model. I did a series of tests to find a sweet-spot for IOPS which happened to be at 160 sessions. The following is a snippet of the AWR report from a 5-minute SLOB run with UPDATE_PCT=0. Since this host has a total of 12 cores I should think 8KB read IOPS of 625,000 per second will impress you. And, yes, these are all db file sequential reads.

load-profile-625K-IOPS

At 52,093 IOPS per CPU core I have to say this is the fastest CPU I’ve ever tested. It takes a phenomenal CPU to handle this rate of db file sequential read payload. So I began to wonder how this would compare to other generations of Xeons. I immediately thought of the Exadata Database Machine data sheets.

Before I share some comparisons I’d like to point out that there was a day when the Exadata data sheets made it clear that IOPS through the Oracle Database buffer cache costs CPU cycles–and, in fact, CPU is often the limiting factor. The following is a snippet from the Exadata Database Machine X2 data sheet that specifically points out that IOPS are generally limited by CPU. I know this. It is, in fact, why I invented SLOB way back in the early 2000s. I’ve never seen an I/O testing kit that can achieve more IOPS per DB CPU than is possible with SLOB.

X2-IOPS-CPU

Oracle stopped using this foot note in the IOPS citations for Exadata Database Machine starting with the X3 generation. I have no idea why they stopped using this correct footnote. Perhaps they thought it was a bit like stating the obvious. I don’t know. Nonetheless, it is true that host CPU is a key limiting factor in a platform’s ability to cycle IOPS through the SGA. As an aside, please refer to this post about calibrate_io for more information about the processor ramifications of SGA versus PGA IOPS.

So, in spite of the fact that Oracle has stopped stating the limiting nature of host CPU on IOPS, I will simply assert the fact in this blog post. Quote me on this:

Everything is a CPU problem

And cycling IOPS through the Oracle SGA is a poster child for my quotable quote.

I think the best way to make my point is to simply take the data from the Exadata Database Machine data sheets and put it in a table that has a row for my E5-2643v3 results as well. Pictures speak thousands of words. And here you go:

Exadata-compare-12c-HSW_EP

AWR Report

If you’d like to read the full AWR report from the E5-2643v3 SLOB test that achieved 625,000 IOPS please click on the following link: AWR (click here).

References

X2 data sheet
X3 data sheet
X4 data sheet
X5 data sheet
X6 data sheet

 


Filed under: oracle

Yes, You Must Use CALIBRATE_IO. No, You Mustn’t Use It To Test Storage Performance.

Fri, 2016-06-03 00:08

I occasionally get questions from customers and colleagues about performance expectations for the Oracle Database procedure called calibrate_io on XtremIO storage. This procedure must be executed in order to update the data dictionary. I assert, however, that it shouldn’t be used to measure platform suitability for Oracle Database physical I/O. The main reason I say this is because calibrate_io is a black box, as it were.

The procedure is, indeed, documented so it can’t possibly be a “black box”, right? Well, consider the fact that the following eight words are the technical detail provided in the Oracle documentation regarding what calibrate_io does:

This procedure calibrates the I/O capabilities of storage.

OK, I admit it. I’m being too harsh. There is also this section of the Oracle documentation that says a few more words about what this procedure does but not enough to make it useful as a platform suitability testing tool.

A Necessary Evil?

Yes, you must run calibrate_io. The measurements gleaned by calibrate_io are used by the query processing runtime (specifically involving Auto DOP functionality). The way I think of it is similar to how I think of gathering statistics for CBO. Gathering statistics generates I/O but I don’t care about the I/O it generates. I only care that CBO might have half a chance of generating a reasonable query plan given a complex SQL statement, schema and the nature of the data contained in the tables. So yes, calibrate_io generates I/O—and this, like I/O generated when gathering statistics, is I/O I do not care about. But why?

Here are some facts about the I/O generated by calibrate_io:

  • The I/O is 100% read
  • The reads are asynchronous
  • The reads are buffered in the process heap (not shared buffers in the SGA)
  • The code doesn’t even peek into the contents of the blocks being read!
  • There is limited control over what tablespaces are accessed for the I/O
  • The results are not predictable
  • The results are not repeatable
My Criticisms

Having provided the above list of calibrate_io characteristics, I feel compelled to elaborate.

About Asynchronous I/O

My main issue with calibrate_io is it performs single-block random reads with asynchronous I/O calls buffered in the process heap. This type of I/O has nothing in common with the main reason random single-block I/O is performed by Oracle Database. The vast majority of single-block random I/O is known as db file sequential read—which is buffered in the SGA and is synchronous I/O. The wait event is called db file sequential read because each synchronous call to the operating system is made sequentially, one after the other by foreground processes. But there is more to SGA-buffered reads than just I/O.

About Server Metadata and Mutual Exclusion

Wrapped up in SGA-buffered I/O is all the necessary overhead of shared-cache management. Oracle can’t just plop a block of data from disk in the SGA and expect that other processes will be able to locate it later. When a process is reading a block into the SGA buffer cache it has to navigate spinlocks for the protected cache contents metadata known as cache buffers chains. Cache buffers chains tracks what blocks are in the buffer cache by their on-disk address.  Buffer caches, like that in the SGA, also need to track the age of buffers. Oracle processes can’t just use any shared buffer. Oracle maintains buffer age in metadata known as cache buffers lru—which is also spinlock-protected metadata.

All of this talk about server metadata means that as the rate of SGA buffer cache block replacement increases—with newly-read blocks from storage—there is also increased pressure on these spinlocks. In other words, faster storage means more pressure on CPU. Scaling spinlocks is a huge CPU problem. It always has been—and even more so on NUMA systems. Testing I/O performance without also involving these critical CPU-intensive code paths provides false comfort when trying to determine platform suitability for Oracle Database.

Since applications to not drive random single-block asynchronous reads in Oracle Database, why measure it? I say don’t! Yes, execute calibrate_io, for reasons related to Auto DOP functionality, but not for a relevant reading of storage subsystem performance.

About User Data

This is one that surprises me quite frequently. It astounds me how quick some folks are to dismiss the importance of test tools that access user data. Say what?  Yes, I routinely point out that neither calibrate_io nor Orion access the data that is being read from storage. All Orion and calibrate_io do is perform the I/O and let the data in the buffer remain untouched.  It always seems strange to me when folks dismiss the relevance of this fact. Is it not database technology we are talking about here? Databases store your data. When you test platform suitability for Oracle Database I hold fast that it is best to 1) use Oracle Database (thus an actual SQL-driven toolkit as opposed to an external kit like Orion or fio or vdbench or any other such tool) and 2) that the test kit access rows of data in the blocks! I’m just that way.

Of course SLOB (and other SQL-driven test kits such as Swingbench do indeed access rows of data). Swingbench handily tests Oracle Database transaction capabilities and SLOB uses SQL to perform maximum I/O per host CPU cycle. Different test kits for different testing.

A Look At Some Testing Results

The first thing about calibrate_io I’ll discuss in this section is how the user is given no control or insight into what data segments are being accessed. Consider the following screenshot which shows:

  1. Use of the calibrate.sql script found under the misc directory in the SLOB kit (SLOB/misc/calibrate.sql) to achieve 371,010 peak IOPS and zero latency. This particular test was executed with a Linux host attached to an XtremIO array. Um, no, the actual latencies are not zero.
  2. I then created a 1TB tablespace. What is not seen in the screenshot is that all the tablespaces in this database are stored in an ASM disk group consisting of 4 XtremIO volumes. So the tablespace called FOO resides in the same ASM disk group. The ASM disk group uses external redundancy.
  3. After adding a 1TB tablespace to the database I once again executed calibrate_io and found that the IOPS increased 13% and latencies remained at zero. Um, no, the actual latencies are not zero!
  4. I then offlined the tablespace called FOO and executed calibrate_io to find that that IOPS fell back to within 1% of the first sample.
  5. Finally, I onlined the tablespace called FOO and the IOPS came back to within 1% of the original sample that included the FOO tablespace.
A Black Box

My objections to this result is calibrate_io is a black box. I’m left with no way to understand why adding a 1TB tablespace improved IOPS. After all, the tablespace was created in the same ASM disk group consisting of block devices provisioned from an all-flash array (XtremIO). There is simply no storage-related reason for the test result to improve as it did.

 

calibrate_after_file_add

More IOPS, More Questions. I Prefer Answers.

I decided to spend some time taking a closer look at calibrate_io but since I wanted more performance capability I moved my testing to an XtremIO array with 4 X-Bricks and used a 2-Socket Xeon E5-2699v3 (HSW-EP 2s36c72t) server to drive the I/O.

The following screenshot shows the result of calibrate_io. This test configuration yielded 572,145 IOPS and, again, zero latency. Um, no, the real latency is not zero. The latencies are sub-millisecond though. The screen shot also shows the commands in the SLOB/misc/calibrate.sql file. The first two arguments to DBMS_RESOURCE_MANAGER.CALIBRATE_IO are “in” parameters. The value seen for parameter 2 is not the default. The next section of this blog post shows a variety of testing with varying values assigned to these parameters.

calibrate_572K-iops

 

As per the documentation, the first parameter to calibrate_io is “approximate number of physical disks” being tested and the second parameter is “the maximum tolerable latency in milliseconds” for the single-block I/O.

table-of-calibrate_io_tests

As the table above shows I varied the “approximate number of physical disks” from 1 to 10,000 and the “maximum tolerable latency” from 10 to 20 and then 100. For each test I measured the elapsed time.

The results show us that the test requires twice the elapsed time with 1 approximate physical disk as it does for with 10,000 approximate physical disks. This is a nonsensical result but without any documentation on what calibrate_io actually does we are simply left scratching our heads. Another oddity is that with 10,000 approximate disks the throughput in megabytes per second is reduced by nearly 40% and that is without regard for the “tolerable latency” value. This is clearly a self-imposed limited within calibrate_io but why is the big question.

I’ll leave you, the reader, to draw your own conclusions about the data in the table. However, I use the set of results with “tolerable latency” set to 20 as validation for one of my indictments above. I stated calibrate_io is not predictable. Simply look at the set of results in the 20 “latency” parameter case and you too will conclude calibrate_io is not predictable.

So How Does CALIBRATE_IO Compare To SLOB?

I get this question quite frequently. Jokingly I say it compares in much the same way a chicken compares to a snake. They both lay eggs. Well, I should say they both perform I/O.

I wrote a few words above about how calibrate_io uses asynchronous I/O calls to test single-block random reads. I also have pointed out that SLOB performs the more correct synchronous single block reads. There is, however, an advanced testing technique many SLOB users employ to test PGA reads with SLOB as opposed to the typical SLOB reads into the SGA. What’s the difference? Well, revisit the section above where I discuss the server metadata management overhead related to reading blocks into the SGA. If you tweak SLOB to perform full scans you will test the flow of data through the PGA and thus the effect of eliminating all the shared-cache overhead. The difference is dramatic because, after all, “everything is a CPU problem.”

In a subsequent blog post I’ll give more details on how to configure SLOB for direct path with single-block reads!

To close out this blog entry I will show a table of test results comparing some key time model data. I collected AWR reports when calibrate_io was running as well as SLOB with direct path reads and then again with the default SLOB with SGA reads. Notice how the direct path SLOB increased IOPS by 19% just because blocks flowed through the PGA as opposed to the SGA. Remember, both of the SLOB results are 100% single-block reads. The only difference is the cache management overhead is removed. This is clearly seen by the difference in DB CPU. When performing the lightweight PGA reads the host was able to drive 29,884 IOPS per DB CPU but the proper SLOB results (SGA buffered) shows the host could only drive 19,306 IOPS per DB CPU. Remember DB CPU represents processor threads utilization on a threaded-processor. These results are from a 2s36c72t (HSW-EP) so these figures could also be stated as per DB CPU or per CPU thread.

If you are testing platforms suitability for Oracle it’s best to not use a test kit that is artificially lightweight. Your OLTP/ERP application uses the SGA so test that!

The table also shows that calibrate_io achieved the highest IOPS but I don’t care one bit about that!

tale-of-the-calibrate-versus-slob-tape

AWR Reports

I’d like to offer the following links to the full AWR reports summarized in the above table:

Additional Reading Summary

Use calibrate_io. Just don’t use it to test platform suitability for Oracle Database.


Filed under: oracle

Is SLOB AWR Generation Really, Really, Really Slow on Oracle Database 11.2.0.4? Yes, Unless…

Tue, 2016-03-08 16:36

If you are testing SLOB against 11.2.0.4 and find that the AWR report generation phase of runit.sh is taking an inordinate amount of time (e.g., more than 10 seconds) then please be aware that, in the SLOB/awr subdirectory, there is a remedy script rightly called 11204-awr-stall-fix.sql.

Simply execute this script when connected to the instance with sysdba privilege and the problem will be solved. 

11.2.0.4-awr-stall-fix.sql

 


Filed under: oracle Tagged: 11.2.0.4, Automatic Workload Repository, AWR, SLOB, SLOB Testing

Reblogged: Providing A Persistent Data Volume to EMC XtremIO Using ClusterHQ Flocker, Docker And Marathon

Thu, 2016-02-25 15:53

I don’t reblog very often–if ever. However, this blog post from EMC’s Itzik Reich is a jewel. If you are like everyone else in IT and are starting to take an interest in Docker I recommend viewing Itzik’s post!

Itzikr's Blog

Hi,

Containers are huge, that’s not a secret to anyone in the IT industry, customers are testing the waters and looking for many ways to utilize containers technologies, it is also not a secret that the leading vendor in this technology is docker.

But docker itself isn’t perfect yet, while it’s as trendy as trendy can get, there are many ways around the docker runtime to provide cluster management etc’..

It all started with a customer request some weeks ago, their request was “can you show us how do you integrate with docker, Marathon (to provide containers orchestration) and ClusterHQ Flocker to provide a persistent data volume..sounds easy right? J

Wait a second, isn’t containers technologies supposed to be completely lossless, designed to fail and do not need any persistent data that will survive a failure in case a container dies??

Well, that’s exactly where customers are asking things that…

View original post 323 more words


Filed under: oracle