Kerry Osborne's Oracle Blog

Wow! – I was stunned a few days ago by Exadata’s Storage Indexes. I was doing a little testing to see what could be offloaded and what couldn’t (more on that later). I have a 384 million row table I was using on our Exadata Quarter Rack test system. A single threaded full scan with no where clause on the table takes about 24 seconds (ho hum – it’s amazing how quickly we become numbed to the outstanding performance ). So imagine my surprise when I decided to check and see how many nulls I had in a column and the result came back in .07 seconds. Wow! I thought it was a bug! Turns out it was the Storage Indexes. Alright already, I’ll show you some output from the system (by the way, as usual I used a couple of scripts: fsx.sql and mystats.sql):

SYS@LABRAT1> select count(*) from kso.skew3;

  COUNT(*)
----------
 384000048

1 row selected.

Elapsed: 00:00:24.06
SYS@LABRAT1> /

  COUNT(*)
----------
 384000048

1 row selected.

Elapsed: 00:00:23.94

SYS@LABRAT1> set timing off
SYS@LABRAT1> @mystats
Enter value for name: %storage%

NAME                                                                             VALUE
---------------------------------------------------------------------- ---------------
cell physical IO bytes saved by storage index                                        0

1 row selected.

SYS@LABRAT1> set timing on
SYS@LABRAT1> select count(*) from kso.skew3 where col1 is null;

  COUNT(*)
----------
        12

1 row selected.

Elapsed: 00:00:00.07
SYS@LABRAT1> set timing off
SYS@LABRAT1> @fsx
Enter value for sql_text: select count(*) from kso.skew3 where col1 is null
Enter value for sql_id: 
Enter value for inst_id: 


 INST SQL_ID         CHILD  PLAN_HASH      EXECS     AVG_ETIME      AVG_LIO    AVG_PIO AVG_PX OFFLOADABLE IO_SAVED_% SQL_TEXT
----- ------------- ------ ---------- ---------- ------------- ------------ ---------- ------ ----------- ---------- ----------------------------------------
    1 0u1q4b7puqz6g      0 2684249835          5           .09    1,956,226  1,956,219      0 Yes             100.00 select count(*) from kso.skew3 where col

1 row selected.

SYS@LABRAT1> @mystats
Enter value for name: %storage%

NAME                                                                             VALUE
---------------------------------------------------------------------- ---------------
cell physical IO bytes saved by storage index                              16012763136

1 row selected.

So apparently Storage Indexes are NULL aware. Very cool! This may have repercussions regarding design and implementation decisions. There are systems that don’t use NULLs in order to insure that they can access records via B-Tree indexes (which as you’re aware do not store NULLs). SAP for example uses a single space character instead of NULLs.

Continue reading ‘Oracle Exadata – Storage Indexes’ »

Over the years Oracle has added many enhancements in order to allow individual SQL statements to take full advantage of multiprocessor computers. A few months ago Cary Millsap did a talk where he recalled the presentation Larry Ellison did when Oracle first announced the Parallel Query feature. During Larry’s demo he had a multiprocessor computer all to himself. I don’t remember how many processors it had, but I remember he had some kind of graphic showing individual CPU utilization on one screen while he fired up a parallel query on another screen. The monitoring screen lit up like a Christmas tree. Every one of the CPU’s was pegged during his demo. When Cary was telling the story he said that he had wondered at the time what would have happened if there had been other users on the system during the demo. Their experience would probably not have been a good one. I remember having the exact same thought.

Oracle’s parallel capabilities have been a great gift but they have also been a curse because controlling the beast in an environment where there are multiple users trying to share the resources is pretty difficult. There have been many attempts at coming up with a reasonable way of throttling big parallel statements along the way. But to date, I think this technology has only been used effectively in batch processing environments and large data warehouses where consuming the whole machine’s resources is acceptable due to the relatively low degree of concurrency required by those environments.

So why did I mention Exadata in the title of this post? Well I believe that one of the most promising aspects of Exadata is it’s potential with regard to running mixed work loads (OLTP and DW) on the same database, without crippling one or the other. In order to do that, Oracle needs some mechanism to separate the workloads. Resource Manager is an option in this area, but it doesn’t go far enough in controlling throughput on parallel queries. This new queuing mechanism should be a great help in that regard. So let’s review the options:

Parallel Adaptive Multi User (the old way)
This ability to automatically downgrade the degree of parallelism based on what’s happening on the system when a query kicks off is actually a powerful mechanism and is the best approach we’ve had prior to 11g Release 2. The downside of this approach is that parallelized statements can have wildly varying execution times. As you can imagine, a statement that gets 32 slaves one time and then gets downgraded to serial execution the next time will probably not make the user very happy. The argument for this type of approach is that stuff is going to run slower if the system is busy regardless of what you do. And that users expect it to run slower when the system is busy. The first part of that statement may be true but I don’t believe the second part is (at least in most cases). The bigger problem with the downgrade mechanism though is that the the decision about how many slaves to use is based on a single point in time (the point when the parallel statement starts). And once the degree of parallelism (DOP) is set for a statement it can not be changed. That execution of the statement will run with the number of slaves it got to start with, even if additional resources become available. So consider the statement that takes a minute with 32 slaves that gets downgraded to serial due to a momentarily high load. And say that 10 seconds after it starts the system load drops back to more normal levels. Unfortunately, the serialized statement will continue to run for nearly 30 minutes with it’s single process, even though on average the system is not busier than usual.

Parallel Queuing (the new way)
Now let’s compare that with the new mechanism introduced in 11gR2 that allows parallel statements to be queued in a First In – First Out fashion. This mechanism separates (presumably) long running parallel queries from the rest of the workload. The mechanics are pretty simple. Turn the feature on. Set a target number of parallel slaves (parallel_servers_target). Run stuff. If a statement tries to start that requires exceeding the target, it will be queued until the required number of slaves become available.

The Parallel Queuing feature is controlled by a hidden parameter called “_parallel_statement_queuing”. A value of TRUE turns it on and FALSE turns it off. FALSE is the default by the way. This parameter is not documented but is set automatically when the PARALLEL_DEGREE_POLICY parameter is set to AUTO. Unfortunately, PARALLEL_DEGREE_POLICY is one of those parameters that controls more than one thing. When set to AUTO it also turns on Automatic DOP calculation. This feature calculates a DOP for each statement regardless of whether any objects have been decorated with a parallel setting. The result is that all kinds of statements are run in parallel, even if no objects have been specifically defined with a parallel degree setting. This is truly automatic parallel processing because the database decides what to run in parallel and with how many slaves. On top of that, by default, the slaves may be spread across multiple nodes in a RAC database (this can be disabled by setting PARALLEL_FORCE_LOCAL to TRUE). Finally, AUTO is supposed to enable “In Memory Parallel Query”. This poorly named feature refers to 11gR2’s ability to make use of the SGA for parallel query, as opposed to using direct reads exclusively. Note: I haven’t actually seen this kick in yet, which is probably good, since Exadata Offloading depends on direct reads. I haven’t seen it kick in on non-Exadata databases either though.

Unfortunately this combination of features is a little like the wild west with things running in parallel all over the place. But the ability to queue parallel statements does provide some semblance of order. And to be fair, there are a number of parameters that can be set to control how the calculations are performed. Anyway, here’s a brief synopsis of parameter changes caused by the various settings PARALLEL_DEGREE_POLICY.

Continue reading ‘Exadata and Parallel Queuing’ »

The “Secret Sauce” for Exadata is its ability to offload processing to the storage tier. Offloading and Smart Scan are two terms that are used somewhat interchangeably. Offloading is a more generic term that means doing work at the storage tier that would otherwise have to be done on the database tier (this can include work that is not related to executing queries such as optimization of incremental backups). Smart Scans on the other hand are the access mechanism used to offload query processing tasks. For example, storage servers can apply predicate filters at the storage layer, instead of shipping every possible block back to the database server(s). Another thing that happens with Smart Scans is that the volume of data returned can be further reduced by column projection (i.e. if you only select 1 column from a 100 column table, there is no need to return the other 99 columns). Offloading is geared to long running queries that access a large amount of data. Offloading only works if Oracle decides to use its direct path read mechanism. Direct path reads have traditionally been done by parallel query slaves, but can also be done by serial queries. In fact, as of 11g, Oracle has changed the decision making process resulting in more aggressive use of serial direct path reads. I’ve seen this feature described both as “serial direct path reads” and “adaptive direct path reads”.

I’ll digress here a bit to discuss this feature since direct path reads are critical to Exadata Offloading. Direct path reads do not load blocks into Oracle’s buffer cache. Instead, the data is returned directly to the PGA of the process requesting the data. This means that the data does not have to be in Oracle block format. That means no 8K block that is only partially filled, that may only have a record or two that you’re interested in, containing every column including ones you don’t want, and with additional header information – needs to be shipped back up from the storage layer. Instead, a much more compact result set containing only the columns requested and hopefully only the rows you need are returned. As I said, direct path reads are traditionally used by parallel query slaves. They are also used in a few other instances such as LOB access and sorts that spill over into TEMP. So the ability to use direct path reads is very important to the Exadata platform and thus the changes to the make them more attractive in 11g. Here are a few links to info on the subject of serial direct path reads:

1. Doug Burns has a good post on 11g serial direct path reads.
2. Alex Fatkulin has a very good post on some of the factors controlling adaptive direct path reads.
3. There is a note on MOS (793845.1) on changes in 11g in the heuristics to choose between direct path reads or reads through the buffer cache.
4. You may also find MOS note (50415.1) on misleading nature of “direct path read” wait events of interest.

Also be aware that direct path reads are only available for full scans (tables or indexes). So any statement that uses an index range scan to get to a row in a table via a rowid will not use this mechanism. Also keep in mind that direct path requires extra processing to ensure that all blocks on disk are current – (i.e. an object level check point), so frequently modified tables will suffer some overhead before direct path reads can be initiated.

I must say that I think the changes to the heuristics in 11g may be a little on the aggressive side for non-Exadata platforms (the changes may well be driven by Exadata). And by the way, serial direct path reads are not always faster than the normal reads that go through the buffer cache. Dion Cho has a good post on a performance problem due to serial direct path reads kicking in on one node of an 11g RAC environment (not Exadata). The node doing the direct path reads was running the query much slower than the node using the normal buffer cache reads. He also has a post on turning off serial direct path reads.

But enough about the direct path reads stuff, on to the Offloading. One of the first things I wanted to know when I got my first look at a system running on Exadata was whether a particular query was eligible for offloading and if so, how much of the expected i/o was saved by the feature. So of course I wrote a little script to show me that. Turns out there is plenty of info in V$SQL to see what’s going on. I called the script fsx.sql (short for Find_Sql_eXadata). Here’s a little demo:

Continue reading ‘Exadata Offload – The Secret Sauce’ »

Well our new Exadata showed up this week. We had a pretty nice lab environment already. A bunch of Dell’s, some IBM’s, several Sun’s. We have a couple of EMC Sans as well (we actually threw away a whole EMC rack to make room for the Exadata). And of course we have every version of Oracle from 8i to 11gR2. It’s a good learning environment. It also let’s us try things when clients have a specific set of versions that we want to mimic. So now we have an Exadata V2 as well. We’ve had the delivery date on the calendar for several weeks. For some reason it reminded of the Weird Al Yankovik song Weasel Stomping Day.

It’s probably a sad reflection on how geeky we are that everyone is running around all excited like it’s Christmas or something.

Here’s a few pictures:

It’s a really fast machine by the way. In fact, we had trouble keeping up with it from the moment we got it off the truck.

Really fast, and slippery. Well in a couple of days we can actually turn it on (we’re supposed to let it acclimate to our environment). On Friday afternoon we’re going to have a happy hour to celebrate our newest edition. Wade calls it a sip and see. We’ll probably take a few pictures of ourselves with the little bundle of joy and sing a chorus of a festive Weird Al song, or maybe two. Come on by if you’re in the neighborhood!

One of the things I didn’t really talk about in my first post on Exadata was the flash cache component of the storage servers. They are a key component of the “OLTP” claims that Oracle is making for the platform. So let’s talk about the hardware first. The storage servers have 4 of the Sun Flash Accelerator F20 PCIe cards. These cards hold 96G each for a total of 384G on each storage server. That’s well over a terabyte on the smallest quarter rack configuration. Here’s what they look like:

Note that they are only installed in the storage servers and not in the database servers. The cards are usually configured exclusively as Flash Cache, but can optionally have a portion defined as a “ram disk”.

Oracle has a White Paper here:

Exadata Smart Flash Cache and the Sun Oracle Database Machine

This white paper was published in late 2009 and it is specific to V2. It has some good information and is well worth reading. One of the comments I found interesting was the discussion of carving a piece of the Flash Cache out as a “disk”. Here’s the quote:

These high-performance logical flash disks can be used to store frequently accessed data. To use them requires advance planning to ensure adequate space is reserved for the tablespaces stored on them. In addition, backup of the data on the flash disks must be done in case media recovery is required, just as it would be done for data stored on conventional disks. This option is primarily useful for highly write intensive workloads where the disk write rate is higher than the disks can keep up with.

Do not confuse the use of these cards in the storage server with the new 11gR2 feature “Database Flash Cache”. That feature allows an extended SGA (level 2) cache to be created on a database server (if you are using Solaris or Oracle Enterprise Linux) and has nothing to do with the Exadata Smart Flash Cache which resides on the Exadata storage servers. Think of the Database Flash Cache as an extended SGA and the Exadata Smart Flash Cache as large “smart” disk cache. I say smart because it implements some of the same type of Oracle cache management features as the SGA.

Kevin Closson has a couple of good posts outlining the differences between Database Flash Cache and Exadata Smart Flash Cache here:

Pardon Me, Where Is That Flash Cache? Part I.
Pardon Me, Where Is That Flash Cache? Part II.

Note also that Exadata Smart Flash Cache does not affect writes (i.e. it is not a write cache).

So how do we see what’s going on with the Exadata Flash Cache? Well there are a couple of ways.

1. We can use the cellcli utility on the storage servers themselves.
2. We can look in v$sesstat (one of the best ways to do that is with Tanel Poder’s snapper script by the way).

Here’s a little output from the system showing method 1 (cellcli):

[root@dm01cel01 ~]# cellcli
CellCLI: Release 11.2.1.2.3 - Production on Fri Apr 30 16:09:29 CDT 2010

Copyright (c) 2007, 2009, Oracle.  All rights reserved.
Cell Efficiency Ratio: 38M

CellCLI> LIST METRICCURRENT WHERE objectType = 'FLASHCACHE'
         FC_BYKEEP_OVERWR                FLASHCACHE      0.0 MB
         FC_BYKEEP_OVERWR_SEC            FLASHCACHE      0.0 MB/sec
         FC_BYKEEP_USED                  FLASHCACHE      300.6 MB
         FC_BY_USED                      FLASHCACHE      135,533.7 MB
         FC_IO_BYKEEP_R                  FLASHCACHE      10,399.4 MB
         FC_IO_BYKEEP_R_SEC              FLASHCACHE      0.0 MB/sec
         FC_IO_BYKEEP_W                  FLASHCACHE      6,378.3 MB
         FC_IO_BYKEEP_W_SEC              FLASHCACHE      0.0 MB/sec
         FC_IO_BY_R                      FLASHCACHE      480,628.3 MB
         FC_IO_BY_R_MISS                 FLASHCACHE      55,142.4 MB
         FC_IO_BY_R_MISS_SEC             FLASHCACHE      0.0 MB/sec
         FC_IO_BY_R_SEC                  FLASHCACHE      0.1 MB/sec
         FC_IO_BY_R_SKIP                 FLASHCACHE      1,448,220.2 MB
         FC_IO_BY_R_SKIP_SEC             FLASHCACHE      12.8 MB/sec
         FC_IO_BY_W                      FLASHCACHE      178,761.9 MB
         FC_IO_BY_W_SEC                  FLASHCACHE      0.1 MB/sec
         FC_IO_ERRS                      FLASHCACHE      0
         FC_IO_RQKEEP_R                  FLASHCACHE      1051647 IO requests
         FC_IO_RQKEEP_R_MISS             FLASHCACHE      291829 IO requests
         FC_IO_RQKEEP_R_MISS_SEC         FLASHCACHE      0.0 IO/sec
         FC_IO_RQKEEP_R_SEC              FLASHCACHE      0.0 IO/sec
         FC_IO_RQKEEP_R_SKIP             FLASHCACHE      0 IO requests
         FC_IO_RQKEEP_R_SKIP_SEC         FLASHCACHE      0.0 IO/sec
         FC_IO_RQKEEP_W                  FLASHCACHE      176405 IO requests
         FC_IO_RQKEEP_W_SEC              FLASHCACHE      0.0 IO/sec
         FC_IO_RQ_R                      FLASHCACHE      21095663 IO requests
         FC_IO_RQ_R_MISS                 FLASHCACHE      1574404 IO requests
         FC_IO_RQ_R_MISS_SEC             FLASHCACHE      0.6 IO/sec
         FC_IO_RQ_R_SEC                  FLASHCACHE      1.6 IO/sec
         FC_IO_RQ_R_SKIP                 FLASHCACHE      4879720 IO requests
         FC_IO_RQ_R_SKIP_SEC             FLASHCACHE      26.8 IO/sec
         FC_IO_RQ_W                      FLASHCACHE      5665344 IO requests
         FC_IO_RQ_W_SEC                  FLASHCACHE      2.9 IO/sec

Continue reading ‘Oracle Exadata V2 – Flash Cache’ »

Well I’ve been holed up playing with an Exadata V2 machine for the past several weeks. Wow. Very interesting technology.

I must say that I believe the concept of offloading SQL processing to the storage layer is a game changer and I wouldn’t be at all surprised to see this as a standard feature a few years from now. What that means for other storage vendors is unclear at this point. So for this first post on the topic let me just describe the configuration (and some potential upgrades).

The basic architecture consists of a set of database severs and a set of storage servers.

Database Servers:

• Sun x4170 (1RU 64x server)
• 2 – Quad-core Intel Xeon E5540 2.53GHz processors
• 72G Ram (18x4G Dimms – max of 144G using 8G DIMMs)
• Dual-Port QDR InfiniBand Host Channel Adapter
• HBA with 512MB Battery Backed Write Cache (only for internal disks???)
• 4 – 146G internal drives (SAS 10,000 RPM)
• dual hot swappable power supplies
• no spare/empty slots!

Here’s what the Database Servers look like:

Storage Servers:

• Sun x4275 (2RU 64x server)
• 2 – Quad-core Intel Xeon E5540 (2.53GHz) processors
• 24G Ram
• Dual-Port QDR InfiniBand Host Channel Adapter
• HBA with 512MB Battery Backed Write Cache (only for internal disks???)
• dual hot swappable power supplies
• 4 – 96G Sun Flash PCIe Cards (total of 384 GB)
• 12 – 600 GB 15,000 RPM SAS or 2 TB 7,200 RPM SATA

Continue reading ‘What is Exadata?’ »