Co-author: Lasse Nedergaard

Yesterday, Lasse ran into an issues with a query pattern in the large database that he is responsible for. Based on our conversation, we wrote up this blog and created a repro.

The troublesome query we were debugging executed like this:

1. Find a list of keys values to search for
2. Insert these keys in a temp table – lets call this the SearchFor table
3. Join the temp table to a large table (lets call it BigTable) and retrieve the full row from the large table

Why not use a correlated sub query in step 2? In this case, the customer in question had multiple code paths (including one accepting XML queries) that all needed to pass thousands of key to a final search procedure. They wanted a generic way to pass these key filters to to the final access of BigTable. The schema will make it clearer.

Schema

We created a repro and it turns out to be a great look into the strange world of database statistics. Here is the schema:

CREATE TABLE BigTable
(
  NearKey UNIQUEIDENTIFIER NOT NULL
  , Payload char(200) NOT NULL
  , RowType INT NOT NULL
)

/* Insert rows to simulate large table with 177 being the skewed value */

INSERT INTO Bigtable WITH (TABLOCK) (NearKey, PayLoad, RowType)
SELECT NEWID()
  , REPLICATE(‘X’, 200)
  , CASE WHEN n%100 BETWEEN 0 AND 50 THEN 177 ELSE n%100 END
FROM dbo.fn_nums(1000000)

/* Creates indexes and statistics */

CREATE INDEX IX_NearKey ON BigTable (NearKey)
CREATE STATISTICS STAT_RowType ON BigTable(RowType)

UPDATE STATISTICS BigTable WITH FULLSCAN

/* Insert rows that are NOT in the statistics with key = 199 */

INSERT INTO BigTable (NearKey, PayLoad, RowType)
SELECT NEWID(), ‘New Rows Not stat updated’, 199 /* New type*/
FROM fn_nums (100000)

/* Create generic search table */

CREATE TABLE #SearchFor (NearKey UNIQUEIDENTIFIER)

 

In Production, BigTable is a very large table with billions of rows. Each row in the table has RowType that describes metadata about the row. The metadata is used to populate #SearchFor before accessing BigTable, as this saves significant I/O.

Things to note about BigTable:

• All stats have been fully updated for values that are not RowType = 199. As long as we stay below this value, we have done ALL we can
• There is big skew in the RowType column.
• There are 100K values that are not in the statistics (you can run DBCC SHOW_STATISTICS (BigTable, STAT_RowType) WITH HISTOGRAM to convince yourself of this)

Query Problems

When a user queries this system, the code will first populate #SearchFor with a stored procedure like this:

Typical search procedures in the system look like this:

CREATE PROCEDURE GetDataA
AS

SELECT … INTO #SearchFor

EXEC dbo.ReturnFromBigTable

 

ReturnFromBigTable contains the troublesome statement that is the subject of this blog:

SELECT *
FROM BigTable
JOIN #SearchFor
  ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = <SomeValue>

 

Looks quite innocent, doesn’t it?

To simulate the customer’s workload, let us say we are searching for 1% of the rows in BigTable. We do this by populating #SearchFor.

INSERT INTO #SearchFor
SELECT TOP 10000 NearKey FROM BigTable   
TABLESAMPLE (1 PERCENT)

And now, the fun begins! Lets first use our NearKey search and for a non skewed RowType:

SELECT * FROM BigTable
JOIN #SearchFor
  ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = 99

Plan (only join shown)

Looking at Predicate in properties for the BigTable scan we unsurprisingly see that the predicate is being driven down: [dbo].[BigTable].[RowType]=(99)… Nothing interesting so far. The plan is parallel. Again, not surprising considering that we are scanning accessing 1M rows in BigTable.

… but as Lasse and I observed on his workload, something isn’t quite right here.

Let us run the same query, but with the filter RowType = 99 replaced by RowType = 177. The new plan is:

It is at this point your design intuition should wake you up with that good ol’ WTF feeling.  “Wait a minute you should say”…before (when there were fewer rows coming out of BigTable) we were parallel and hashing the output of BigTable. Now, when there are more (50x more) rows, we are both reversing the hash order, probing into BigTable and hashing SearchFor. Note that all estimates are relatively accurate (within 10%) and statistics in the space we are searching is fully up to date… If we take a fully naïve approach, trusting relational algebra and query optimizers, there is nothing more we can do here.

But hey, it gets worse: What happens when we go outside the boundaries of the statistics. This for example, could happen when the stats are not fully updated – a completely normal situation in a database. Searching with RowType = 199 (that is not in the histogram) and estimating the plan, we get:

The estimate coming out of BigTable is 1 row, the actual is 100K; this is probably going to hurt.

Running the query gives us the deadly:

Congratulations to us, we just committed performance suicide!

Lets just list the execution times of these query variants:

Join Strategy	RowType	CPU time	Elapsed	Logical I/O Reads
Hash BigTable Probe #SearchFor	99	421ms	549ms	31429
Hash #SearchFor Probe BigTable	177	765ms	1293ms	31429+62
Merge Join (sort both)	199	1575ms	2261ms	31429+62

What is the problem here? From the user’s perspective, the query response time will often vary unpredictably – which is a poor experience. But worse, from a system and DBA perspective: the memory consumption will vary a lot depending on which query plan we end up with. This means it becomes hard to extrapolate the total system throughput and resource requirements based on the queries. If you are capitalist inclined in your mindset, let me translate this: At the end of the day, this costs you money!

Why is this happening? It is happening because there are hidden assumptions here that you have not brought into the light. We are trusting technology (in this case the database statistics and optimizer) with solving something that is a data modeling problem.

This is a very good example of something Thomas has been seeing a lot lately: An unwillingness to be specific about your requirements costs you money – a LOT of money. In other words: There is a price tag associated with ignorance. Let us make it clear what we are NOT saying: There is nothing wrong with ignorance: it is hard to know what you don’t know. Sometimes, you simply don’t know a requirement exists. The trouble arrives when denial and lack of curiosity and analysis trumps the observed data.

Fortunately, Lasse is the curious kind and he asked all the right questions and sought knowledge to remedy the situation.

Let’s get pragmatic, and talk about solutions.

The solution

With all the query plans above, which one do we want? In this case, the answer is: “None of them”

Here is an observation about the workload we discovered as we were analysing the solution: It is reasonable to assume ask the user to accept that response time is proportional with the number of results returned. From a system perspective, we also want the memory consumption and CPU to be proportional to the returned result size. These are questions you have to ask yourself as system designer and discussion you can have with users.

This is where computer science comes into the equation. We know:

1. That we can search a B-tree of n rows in lg(n) time (which is as good as constant)
2. That the the returned result are capped by the number of rows (let us call this R) in #SearchFor
3. That we are willing to accept that the runtime and resource use is proportional with number of rows returned
4. That we cant live with unpredictable behavior

The optimizer KNOWS about the search time and the cap (1+2). But the optimizer does NOT know that we are willing to declare certain properties that narrow the search spare (3) and that we want predictability (4). Our "ignorance” has a price, and the optimizer makes a choice that might be optimal, but which unfortunately is also unpredictable.

This is a great case for hinting. And we can do it like this:

SELECT *
FROM #SearchFor
INNER LOOP JOIN BigTable
  ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = 177

Note something important here: The order of the joined tables matter when you use the LOOP hint (which is why we flipped them above). The above hint BOTH forces the join order and the join strategy.

An alternative, which may be more generally applicable is this rewrite:

SELECT *
FROM BigTable
INNER JOIN #SearchFor WITH (FORCESEEK)
ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = 177
OPTION (LOOP JOIN)

The above rewrite does not guarantee join ordering. However, with the FORCESEEK hint, the plan we want is almost guaranteed.

Let us add some new rows to the result:

Join Strategy	RowType	CPU time	Elapsed	Logical I/O Reads
Hash BigTable Probe #SearchFor	99	421ms	549ms	31429
Hash #SearchFor Probe BigTable	177	765ms	1293ms	31429+62
Merge Join (sort both)	199	1575ms	2261ms	31429+62
Hinted LOOP	99	406	718	81259
Hinted LOOP	177	499	964	81259
Hinted LOOP	199	306	718	81259

Above, we can see why the optimizer didn’t want to help us! The number of IOPS required for my hinted plan are much higher than the optimizer generated plans.

The hinted plan shape is:

Obviously, we could replace the non clustered index with a clustered on (which would get rid of the RID lookup) to make this query faster. But in this case, this was not a viable solution for other reasons.

Summary

In this blog, I have shown you an example of database optimizers failing to make the “right choice”. There are several reasons this is happening:

Lack of information: The database needs declarative information about the user’s intent and what we consider the Right Choice™. In the example, the missing information is: “the speed of the returned result should always be proportional with the result size”. We used a hint to control the optimizers behavior, using our knowledge of the physical execution of queries. Other people might have taken more extreme steps and argued for noSQL – perhaps too aggressive a step in this case.

Leaky abstractions: We saw the line between the “logical” and “physical” model break down. We believed that by building the right model, we had supplied enough information to the database. The distinction between physical and logical model has, as Thomas previously argued, always been a pretty useless paradigm. We have to consider the components of the system in a holistic way, but not at the expense of maintaining the overview of individual building blocks.

Optimality vs. Good enough: the optimizers default behavior is to look for “optimal plans”. In the plan search space, even a very small space like this example, plans are sometimes found that look optimal from statistics. Yet, these plans may not be optimal or have the properties we seek – or we may fail to recognize the good plans as we search the space. In our case, we just wanted a plan that was “good enough” and had certain predictable attributes.

False hardware assumptions: The optimizer makes some assumptions about the cost of IOPS which unfortunately do not correlate with the reality of a modern machine. It assumes that optimizing for a low number of Logical IOPS is a Good Thing™. In this light, the plans generated are the right plans. However, the optimizer does not take into account that the I/O system might be fast, and that there is enough RAM to have a likelihood of finding some of the needed data already residing in the buffer pool. To prove this: Recall that we had fully updated statistics available for queries going after RowType = 177. But, running the query on a petty IBM laptop the  loop hinted plan, even on an empty buffer pool, is still faster than running the plan generated by the optimizer.

The simple query we studied here raise some interesting questions about the design tradeoffs that the data modeler and architect is forced to make. In this case, we were fortunate that Lasse was on the alert for these tradeoffs. With this example, we have seen what the potential consequences of “behavior and model ignorance” are and why you need to be alert to it. 

Location: Somewhere over the Atlantic

When designing highly scalable architectures for modern machines, you will often need to do some form of manual parallelism control. Managing this is not always easy, but in this blog I will give you one piece of my toolbox to help you.

Let us walk through an example together, a tiny case study. This is a problem which many of you will be familiar with.

Let us say you have 16 files that you want to load into the same table in your database in an automated manner. The naïve approach will do something like this:

BULK INSERT MyTarget FROM ‘C:\temp\MyFile1′ WITH
  (FIELDTERMINATOR = ‘;’, ROWTERMINATOR = ‘\n’)

BULK INSERT MyTarget FROM ‘C:\temp\MyFile2′ WITH
  (FIELDTERMINATOR = ‘;’, ROWTERMINATOR = ‘\n’)

… etc…

BULK INSERT MyTarget FROM ‘C:\temp\MyFile16′ WITH
  (FIELDTERMINATOR = ‘;’, ROWTERMINATOR = ‘\n’)

Now here is the problem with this approach: it executes one statement at a time. Sequential execution is BAD, you need to stop thinking about the world like that if you want to scale on a modern architecture.

Lets assume we have enough hardware resources (in this case, it would take a blade server and a decent I/O system). What we really want is to run every one of these statements in parallel. Unfortunately, SQL server does not have a command to start up new connection from inside T-SQL… what to do?

Getting to the Command Line

Because you cannot execute more than one command on a single connection at a time, we will need multiple connections to SQL Server and this mean we have to go back to the command line. Let us start by creating a little batch file Worker.Cmd with this content:

REM Worker.Cmd File

CALL SQLCMD –S.\MyServer –q”BULK INSERT MyTarget FROM ‘C:\Temp\MyFile%1’ …EXIT

This allows us to invoke a bulk load for the first file by executing: Worker.Cmd 1

Unfortunately, we still cannot start multiple connections without manually firing up a lot of command prompts. The coders in the audience may at this point reach for their favorites programming language to write a little utility that can spawn multiple copies of an executable.

However, there is a problem with such a home made executable: you cannot generally rely on a server having the necessary runtime libraries. Typical comments might be:

 “No, we don’t have .NET 4.0 here, this is not yet certified by our infrastructure department. Could you recompile it for 1.1 please?”.

“Power Shell is much too fancy for us, what is wrong with running this on Windows 2000?”

Perhaps this customer is just skeptical about letting you run your executable on a server. This may sound silly, but I have seen this happen too many times to make assumptions.

Start to the Rescue

There is a very nice little utility for the good old command prompt that allows you to fire up new processes: START.EXE. This comes with all versions of Windows and it takes any command line executable as input, fires it up in a new thread and returns control back to the caller.

Using start.exe, we can write batch script that fire up multiple copies of the same executable. It looks like this:

REM SpawnMany.cmd

REM Author: Thomas Kejser
REM Purpose: Spawns many copies of the same executable. Useful for running many things in parallel

@ECHO OFF
ECHO Spawning %2 copies of %1

FOR /L %%i IN (1, 1, %2) DO (
    ECHO Spawning thread %%i
    START "Worker%%i" /Min %1 %%i %2
)

Each new process is started in a minimized window and we pass the thread number and the total number of threads to it. Using this little batch script, we can now do this:

  SpawnMany.exe Worker.Exe 16

This starts 16 workers, each with their own thread number assigned. Very useful for running stuff in parallel in a quick and dirty way. For example, I use this to run the TPC-H data generator dbgen.exe highly parallelized.

Notice that I added the EXIT command at the end of the worker.cmd batch. This makes sure that the window closes itself when done executing.

Summary

In this blog, I have shown you how to write a little batch script to fire up multiple threads, from the command line. each doing their own work in parallel. The script is “zero dependency” which makes it ideal for server use and for hacking together quick and dirty parallelism for test scenarios.

I mentioned that SQL server does not have a way to start up new connections from T-SQL. This is not strictly true. Sorry for leading you astray, but I wanted you to see how to do this from the command line first (and go through the pain ). There is a way to hack SQL Server and implement a stored procedure I like to call sp_executesql_async. This will be the subject of a future blog, but since I am heading into a lab for a few weeks, you just have to wait for it.

As we saw in my previous post, the build in hash functions of SQL Server were either expensive with good distribution, or cheap, but with poor distribution. As a breath of fresh air, let us look at a useful magic quadrant:

 

We see that all of the hash functions exposed by HASHBYTES fall into low speed quadrants, caution is advised here. We also see that while CHECKSUM and BINARY_CHECKSUM are challengers, they do not have the spread to fully compete. We are left with modulo as the best possible hash function, but we believe that this algorithm may have trouble executing. The Modulo hash function has some problems (sorry, had to say “problem” instead of “challenge”, I couldn’t bear speaking management bollocks anymore):

• It only works on integer types
• If there is structure in the data (for example, if all values are equal) it will not spread the values equally

The question we are faced with is:

“Could we create a high speed, good spread, hash function that belongs in the leader quadrant and which does not have the limitation of the module functions”

It was this question I sought to answer when I started hacking away at SQLCLR functions. As you read the following, please don’t laugh too hard at me (a small snigger will suffice) – it has been a long time since I hacked away at code.

MurmurHash2 and Davy’s Bit Magic

Google has published a very interesting hash function, MurmurHash, claiming to have the properties I am looking for: It is both fast and has a good spread. The hash algorithm comes in different flavours: MurmurHash2 and MurmurHash3. Each flavour also exists in architecture and word size optimized versions. I decided to implement the x86/x64, 32-bit integer version in C#.

Fortunately, someone already beat me to it. Davy Landman has an implementation of MurmurHash2 – which he has kindly shared under GPL. I took his implementation and wrapped it in a SQLCLR user defined function.

While trying to understand Davy’s code and hacking it to work with SQLCLR, I had a few “aha moments” that I would like to share.

First of all, Davy uses this cute trick to turn 4 bytes from a BYTE[] array into a 32-bit INT:

UInt32 k = (UInt32)(data[currentIndex++] 
                  | data[currentIndex++] << 8
                  | data[currentIndex++] << 16
                  | data[currentIndex++] << 24
            );

 

I thought that was a rather neat trick. To practice my illustration skills, here is what happens:

Please note that left shift (<<) and or (|) operators take INT32, not UINT32 as input. Hence the need to cast the final value to UINT32.

Speaking of integers, in my modification of Davy’s source code you will find:

unchecked
{
    return (SqlInt32)(Int32)h;
}

I need to get form the UINT which is returned by the reference implementation of MurmurHash to the signed integer (SqlInt32) that SQL Server understands. Doing the unchecked cast here is faster than Convert.ToInt32.

MurmurHash3

From the C++ source code of MurmurHash3, and using Davy’s bit trick, it took me only a few hours to get myself a nice implementation of MurmurHash3. I compared this with a C++ implementation done by my colleague Christian Martinez (based directly on the Google source), and we agreed on outputs that exercise all the branches in the code. So I am reasonably confident that my implementation is correct.

Except for the fact that it turns out to be important to use SqlBinary instead of SqlBytes (blogged here), there is not much more to say. Feel free to use the MurmurHash3 for your own implementations (my implementation is GPL)

The source code for the MurmurHash functions is too long to paste in this blog, so I created a new page on my site which you can find by following this link: C# Source code for MurmurHash in SQLCLR.

Before I show you the spread and speed results, let me just talk a little about two very old, but very commonly used, hash functions.

CRC16 and CRC32

The Cyclic Redundancy Check (CRC) family of hashes are, compared to MurmurHash, very simple to implement. They walk through each byte in the input and divide it with a specially selected polynomial. For implementation purposes, the polynomial is simply a constant in the code and the division is done with XOR and shifting (making this a very efficient operation). The remainder of the byte division is fed into the division on the next byte (giving the hash algorithm its name) and so on, until the end of the input .

WikiPedia contains a great reference implementation that I used for my tests. For readability, the source code is again available on a separate page of this blog: CRC16 and CRC32 in C#.

In the naïve approach to CRC, the polynomial division is implemented using an inner loop through all the bits of each byte:

rem = rem ^ data[i];

for (int j = 0; j < 8; j++)
{
  if ((rem & 0×00000001) == 0×00000001)
  {   // if rightmost (least significant) bit is set
     rem = (rem >> 1) ^ Polynomial32; /* Polynomial */
  }
  else
  {
    rem = (rem >> 1);
  }
}

An interesting optimization, shaving off a lot of cycles, is to create a table that allows you to look up the result of this loop for all 256 combinations of 8 bits. My source code contains such  tables, both for CRC32 and CRC16 implementations. Using these tables, I can replace the above code fragment with:

rem = (rem >> 8 ) ^ CRCTable32[(rem & 0xff) ^ data[i]];

…A major optimization

Speed Results

At this point, you may wonder if I did all of this to give you a chance to laugh at my rusty coding skills. What exactly was the point of implementing a new hash function in the first place?

Without further ado, let us look at the results compared to native SQL Server functions:

Isn’t that beautiful? If we don’t need the cryptographic properties of the HASHBYTES functions, we can beat SQL Servers hashes with our own SQLCLR implementation. Please note that I took out MD2 from the above to avoid skewing the results (I have previously shown that it is simply too inefficient).

Now, the bad news: the yellow line above. That represent a “hash function” that returns zero, like this:

public static SqlInt32 NoHash(SqlBinary data)
{      
  return (SqlInt32)0;
}

 

Notice that even such a simple user defined function that does nothing, consumes almost all the same number of CPU cycles as a full hash calculation. This is the overhead of the CLR and throwing data back and forth between managed and unmanaged code. So much for my optimizations in CRC.

Spread Results

We now know that there is at least hope of implementing a faster hash function than HASHBYTES. But how well do these functions spread the input. Running the chi-squared test from my previous blog entry we get These results:

Now THAT is pretty nice isn’t it? We see that good old CRC makes for an excellent hash function for our purpose. MurmurHash does very well too, comparable with SHA, but at under half the CPU cost.

Summary

In this blog, I have shown you how to implement hash functions that are both faster and have better spread than the built in hash functions exposed in SQL Server HASHBYTES. We have also seen that the overhead of SQLCLR is significant for this case, and hence, it would be preferable if non cryptographic hash function were exposed natively in SQL Server.

If you care about such new hash functions for SQL Server, I would suggest you file a Connect item to ask for them and state your purpose for wanting them. “It would be cool” or “Thomas says so” is not enough, provide a business justification. But at least, now you know what to ask for…

I owe Christian Martinez thanks for this blog for directing me to the MurmurHash functions, and for helping me validate the correctness of my C# implementation against his C++ version.

PS: As I was exploring this little implementation, I noticed that one of my heroes, Donald Knuth, has extended The Art of Computer Programming with a new volume: Volume 4, Fascicle 1: Bitwise Tricks & Techniques; Binary Decision Diagrams. That sounds like an interesting read and it is now on my birthday wish list.

Any code optimized for highly concurrent workloads must worry about thread synchronization. SQL Server is no exception, because in a database system, synchronization is one of the core functionalities you rely on the engine to provide (the noSQL crowd may ponder that a bit). In this blog post, I will describe the synchronization primitives, like locks, latches and spinlocks, used by SQL Server to coordinate access to memory regions between threads.

Before I go on, an update: am currently sitting in the train from Stockholm after a hugely successful PASS SQL Rally Nordic conference.

At the conference, I presented my new “Grade of the Steel” deck, which I hope will be available online soon. The deck led to a lot of questions about spinlocks during the party that same evening.

We had Swedish legend E-Type playing and putting on an impressive stage show (including dancing Vikings, as seen in picture to the right). The show left me wondering how to improve my presentations skills to get THAT level of reach to the audience. The conference arrangers had  also done an amazing job at keeping everything running smoothly and I return with many good memories. Thank you guys, you showed the crowd that the Nordic community is very strong.

Well, “here I go again”, being distracted by all the fun. My point was, the questions at the conference led me to write this blog entry, as your (always not so) humble servant. Keep that feedback coming.

Locks and Lock Graphs

Most DBAs and developers are familiar with locking and blocking: Locks in SQL Server are the semaphores that guard the consistency of the database and enforce the high level ACID properties and serialization properties in the database.

You can measure locking activity in several DMV, and there are many articles written on how to diagnose and troubleshoot locking issues. For this blog, suffice to say that the total time the entire engine spends on waiting for locks is measured in sys.dm_os_wait_stats. I would like to make the point that EVERY performance investigation I do on SQL Server stats with querying this view… Learn to use it.

In sys.dm_os_wait_stats, you will find rows that have wait_type beginning with LCK. These are the lock waits, and they have been extensively described elsewhere.

The one thing I would like to add, is that locks have a directed acyclic graph relationship (DAG) with each other and are generally acquired in a certain order. Let us take a hypothetical example of a query locking five rows in a table with two partitions:

Before the query can even start, it needs a database (DB) lock to even run, this lock is acquired by the connection itself and held as long as you are connected to the database.

The first locks acquired by the query itself, are the two schema locks (SCH in shared mode S): one of the table itself, and one for the schema the table lives in. Note that the term “schema” is a bit overloaded here: I mean to say that the lock is acquired both on the table (the table object in the database schema) and on the schema (the schema object in the database schema). The SCH locks prevent the table structure itself from being modified while the query runs. The requirement to hold this SCH lock is why, for example a SWITCH operation, will block a person from reading the table while the switch happens.

When SCH locks are acquired, the query must hold a lock on the table. This prevents other conflicting locks, for example a bulk inserter or a greedy, table escalated X-lock, from conflicting with the query.

Traversing onwards from the table level, we have a special lock level: HOBT, representing a partition (HOBT = Heap Or B-Tree, pronounced: “Hobbit”). Recall that even an unpartitioned table has one partition: the table itself. The HOBT lock is only acquired if the table has been configured to support this lock level. This can be done with this statement:

ALTER TABLE <table> SET (LOCK_ESCALATION = AUTO)

 

… note that the default for tables is to have this lock type disabled and to jump directly to…

Row and page locks: Queries will generally lock the pages it needs access to, though this is mostly a performance optimization used for table scans. Row locks fall into two different categories, RID or Key, depending on whether it is a heap or a B-tree you are locking. Note that you can disable row and page locks by altering the table like this for a B-tree:

ALTER INDEX <index> ON <table>
SET (
    ALLOW_ROW_LOCKS = OFF
  , ALLOW_PAGE_LOCKS = OFF)

 

There are cases where disabling the row and page lock levels make queries go faster, but it can also have a negative effect on concurrency. I hope to blog about those cases, and how to diagnose them, in a later installment.

I said before that locks are acquired in a the order of the directed acyclic graph by the SQL engine. This is not something you need to worry about, the engine handles this for you. But, can you guess why it matters which order locks get acquired?

(Hint: What would happen if someone started at the top at the same time that someone else started from the bottom of the graph?)

Locks in SQL Server are stored in hash tables in the part of the engine called the lock manager. These hash tables consume some memory (but so little we don’t really care about it, so Oracle people, stop bickering this point) and are optimized for highly concurrent access. The lock manager even partitions locks over multiple NUMA nodes to make accessing lock information faster. The mechanism used to partition locks are similar to the SuperLatches described in: Hot It Works: SQL Server SuperLatch’ing / Sub-latches.

Latches

Latches are a special type of lock that is used inside the SQL Server engine to protect shared data structures. They enforce integrity at the physical layer of the database and will be used automatically by SQL Server, irrespective of the transactional semantics you request.

The wait time for latches shows up in sys.dm_os_wait_stats, there are three broad categories: PAGEIOLATCH, PAGELATCH and LATCH

The first two are latches used by the database to protect the integrity of the buffer pool. The PAGEIOLATCH variant is used when data is brought in from disk, while the PAGELATCH is used to coordinate things that are already in memory.

The last latch type, LATCH, is a different beast that is used several places inside the SQL code path to coordinate other memory structures – for example during query execution. The good news is that you can get more details about this latch from the DMV: sys.dm_os_latch_stats. You can think of this view as a way to “zoom in” on latches.

The view contains a latch_class called BUFFER – this is just the aggregate of all the PAGEIOLATCH and PAGELATCH waits, you already have the detail of those, so just ignore it. The rest of the rows in sys.dm_os_latch_stats add up to the LATCH waits.

Spinlocks

Locks and latches have something in common: When a thread in SQL server needs to wait for them, the thread will block and yield the scheduler. In general, it is a good idea to play nice this way and yield the scheduler to another thread if you cannot get any more work done until you acquire access to the structures you want to use. SQL Server also implements user mode scheduling, which takes down the price of a context switch significantly – further boosting scale.

Think about locks and latches with this analogy: You are about to leave a plane, but your luggage is in an overhead compartment in the seats behind you. The best strategy for optimizing throughput, is for you to sit down, wait for the people in the seats behind you to pass and then go open the overhead compartment. This optimizes for the best system throughput and lets you turn off your brain just a little longer.

HOWEVER, sitting down and patiently waiting for other people to get their work done first, is not always the smartest idea. As you leave the plane and move to passport control, you don’t sit down on the floor while you wait for the person in front of you to fumble for his passport in his pocket. The reason you don’t do this, is that you expect him to quickly be done with his fetching; and you remain alert while you wait (and curse). It would be a waste of time and leg muscles to sit down, and the people behind you would get annoyed by you too.

Spinlocks are optimized to protect data structure where you expect the thread (or person) in front of you to get their work done fast! They are typically implemented as busy waiting loops that are optimized for the CPU architecture they run on (we want the loop to be as cheap as possible and we need some low level tricks to get it working). Wikipedia has an excellent article with x86 assembly examples.

SQL Server uses spinlocks in many parts of the code where the assumption of “shortly held locks” can be made, and typically this gives you more throughput and is very efficient.

You can see how many spins are done by querying this view:

SELECT * FROM sys.dm_os_spinlock_stats
ORDER BY backoffs DESC

 

In older version of SQL Server (before 2005) – this view was not available. However, there is another way to get the spinlock information, namely by executing:

DBCC SQLPERF ( spinlockstats )

 

The problem with spinlocks is of course that the assumption about shortly held locks is not always true. Sometimes, the guy in front of you just take too long to get stuff done, and you are stuck with all your attention pinned to him (spinning) and getting more and more annoyed. Interestingly, NUMA systems are typically the machines where this assumption breaks down, because accessing memory on a foreign node can take a significant amount of time. If access to that memory is protected by a spinlock, this means you can burn a lot of CPU in the busy waiting loop and you can get some very interesting side effects.

In general, you should not worry about high values in the spins column of sys.dm_os_spinlock_stats. Spinlocks are after all designed to spin a lot in busy waiting loops. The condition that may cause you to worry that the person in front of you is not clearing the queue fast enough, is when you see high values in the backoffs column. Backoffs are what happen when the thread spinning for a resource gives up and decided to fall asleep. If the backups go up dramatically as you increase throughput, you may have spinning on a memory region that is heavily contended and have to design around it.

There is an excellent article on how to diagnose spinlocks written by SQLCAT members Mike Ruthruff and Ewan Fairweather: Diagnosing and Resolving Spinlock Contention on SQL Server. Among other good stuff, it details how to use Trace Flag 3656 (we all love trace flags) and XEvents (and we love XEvents even more) to troubleshooting spinlock problems (see page 14-16).

Interlocked Operations

I am including this last synchronization primitive for completeness (read on to understand why).

Modern x86/x64 CPUs expose some very interesting instructions that allow you to build what is known as interlock operations. An interlock is a CPU special instruction, or series of CPU instructions, that are guaranteed to perform an atomic operation on an integer. For example, using an interlock you can safely add one to an integer and not have to worry that someone else is trying to add one at the same time. As another example, spinlocks (see above) are typically implemented using the compare and swap operation (CMPXCH in the x86 instruction set) that allow you to compare two values and swap them if they are equal, in a single, atomic instruction.

The reason I am only including this section for completeness, is that you have no way of telling if SQL Server used an interlock to synchronize access to a memory region (though it sometimes does) and no way to influence it. And the reason is that an interlock neither blocks, nor spins the thread – it simply allows the thread to execute as if no other threads were touching the data, because the interlock instruction “knows” that the underlying hardware handles the synchronization. It is the ultimate, lightweight synchronization trick.

You may wonder why SQL Server does not use interlocks everywhere. The big reason is that interlocks are severely limited. They can typically only operate on a a single integer at a time, and only support a small subset of operations on that integer. This is not enough to fully implement all the thread synchronization you need.

If you are interesting in reading more about interlock operations, there is a very nice introduction at this site for computer game programmers: Multithreaded Programming Part 3: Going Lockless. And by the way, I recommend studying the work of computer game programmers – they are the true wizards of computer science.

Related:

• Grade of The Steel Category
• Secret Scale Sauce : Heavy Singleton INSERT spinlocks
• Advanced performance troubleshooting: waits, latches, spinlocks (At Paul Randal’s site)

Hash distributing rows is a wonderful trick that I often apply. It forms one of the foundations for most scale-out architectures. It is therefore natural to ask which hash functions are most efficient, so we may chose intelligently between them.

In this blog post, I will benchmark the build in function in SQL Server. I will focus on answering two questions:

• How fast is the hash function?
• How well does the hash function spread data over a 32-bit integer space

I know there is also the question about how cryptographically safe the function is, but this is not a necessary property for scale-out purposes – hence, that aspect is out of scope for this blog.

For the picky people (my dear soul mates): I will be using 1K = 1024 in the following, so 64K = 65536.

Test data

Because we sometimes find ourselves hashing dimension keys, I will use a very simply table with a nice key range as my test input. Such a key range could for example have been generated by an IDENTITY(1,1) column or a SEQUENCER. Here is the test script that generates the input:

CREATE TABLE CustKey (SK INT NOT NULL)

INSERT INTO CustKey WITH (TABLOCK) (SK)
SELECT n FROM dbo.fn_nums(10000000)

 

(see my previous post for the fn_nums source)

Speed of the Hash function

SQL Server exposes a series of hash functions that can be used to generate a hash based on one or more columns.

The most basic functions are CHECKSUM and BINARY_CHECKSUM. These two functions each take a column as input and outputs a 32-bit integer.

Inside SQL Server, you will also find the HASHBYTES function. This little gem can generate hashes using MD2, MD4, MD5, SHA and SHA1 algorithms. The problem for the purpose of our test is that these function spit out BINARY types, either 128 bits (for SHA) or 160 bits (for MD). However, we can quickly map that value into the integer space by doing a modulo (in SQL: %) with MaxInt (2**31 – 1). Interetingly, Modulo work directly on values of BINARY in SQL Server.

Before I move on the results, I would like share a little trick. When you need to get the raw CPU speed of the hash, you may be tempted to do something like this:

SELECT BINARY_CHECKSUM(SK) AS h
FROM CustKey

 

But this is wrong. Why? Because you are both measuring the time it takes to hash as well as the time it takes to transfer the rows to the client. This is not very useful, on my laptop it takes over two minutes to run the above query.

Instead, you should try to measure the raw time the server needs to calculate the data. One way to achieve this (and this trick can be used on most queries you benchmark) is to wrap the SELECT statement in a MAX function. Like this:

SELECT MAX(h)
FROM (
    SELECT BINARY_CHECKSUM(SK) AS h
    FROM CustKey
    ) noRows
OPTION (MAXDOP 1)

Notice something else interesting, I am forcing parallelism down to 1. I want to get as close to the raw cost of the CPU as possible – so I don’t want to clutter the query with any overhead of parallelism. The query now runs in a little less than 3 seconds on my laptop.

I addition to the above trick, we also need to quantify the time SQL Server spends on just accessing the table. We can get a good approximation of this by measuring the SELECT statement without any hash function. On my laptop, I measured this to be around 3200 ms. In the results below, I have subtracted this runtime so we are only measuring the cost of the hashing.

With all this said, let us have a look at the results:

Some comments on the data:

• Notice that MD2 is very inefficient. I have not studied this algorithm in great detail and would appreciate a specialist commenting here.
• Apart from the MD2 anomaly, the HASHBYTES functions all performan pretty much the same. Using around 1 micro sec per hash value calculation
• BINARY_CHECKSUM and CHECKSUM are MUCH faster, by about an order of magnitude
• Doing a simple modulo that spreads out the data over the integer space is about the same speed as CHECKSUM and BINARY_CHECKSUM
• It is not the cast to NVARCHAR that uses CPU time (HASHBYTES requires NVARCHAR input)
• However, the checksum function do take longer to run when you first cast to NVARCHAR. Costing around 100 ns per INT value

Of course, the HASHBYTES functions also have other characteristics than speed which may be desired (cryptographic utility for example). but if you you want is speed, it looks like CHECKSUM and BINARY_CHECKSUM are faster.

But, let us see how good the functions are at spreading the data.

Spread of the Hash Function

Another desirable characteristic, apart from speed, of a hash function is how well it spreads values over the target bit space. This is useful not only for cryptographic purposes, but also to “bucket” data values into equal sized portions, for example in scale-out MPP architectures.

Often, you will not know in advance how many buckets you eventually want to subdivide the integer space into. You may start out with 32K buckets and divide up the integer interval in 128K (32-bit space divided by 128K = 32K) even sized buckets, each with 128K values in them:

• Bucket0: [MinInt…MinInt + 128K[
• Bucket1: [MinInt + 128K…MinInt + 128K*2[
• Bucket2: [MinInt + 128K*2…MinInt + 128K*3[
• … etc..
• Bucket 32K: [MinInt + 128K*32K…MaxInt]

Perhaps you will later want a further subdivision into 64K buckets, each with 64K hash values.

Wise from the runtimes I measured before, I used only 1M rows with values [1…1M] for this test. However, this is plenty to show some good results, so don’t worry.

To avoid running chi-square testing (See later) of integer value spread all over the 32-bit space, I calculated the hash of each key and I then bucketed them into 64K bucket ranges (each with 64K values, neat isn’t it?).

This table comes in handy for that bucketing

CREATE TABLE ExpectedBuckets 
(
    BucketID INT
  , RangeStart INT
  , RangeEnd INT)

INSERT INTO ExpectedBuckets
SELECT
    n
    ,(n – 1 – POWER(2,15)) * POWER(2,16)
    , (n – POWER(2,15)) * POWER(2,16) – 1
FROM dbo.fn_nums( POWER(2,16) )

A sample output form this table:

I can calculate the hash values and put them into a table like this:

CREATE TABLE HashResult
(
  ObservationID INT IDENTITY(1,1)
, HashValue_i INT
)

CREATE CLUSTERED INDEX CIX ON HashResult (HashValue_i)

 

And finally, I can join them all up and count the values in each bucket:

SELECT
      BucketID
    , CAST(ISNULL(COUNT(HashValue_i), 0) AS FLOAT) AS Observed
FROM ExpectedBuckets EB
LEFT JOIN HashResult
ON HashValue_i BETWEEN EB.RangeStart AND EB.RangeEnd
GROUP BY BucketID

 

We now have a nice count of the content in hash bucket, it looks like this:

How do we test if this is a good result or not? It seems that min/max, averages and standard deviations don’t really suffice here (sorry modern MBA types, I realise I am going to loose you now ).

We need to do some hypothesis testing. The hypothesis I want to test (aka: the NULL hypothesis – all resemblance to relational algebra is purely coincidental).

“My chosen hash function evenly distributes the integers into the 64K hash buckets”

Interestingly, this hypothesis makes a prediction (has to, if not we cannot falsify it): namely that in each bucket filled by hashing the 1M rows, we should expect to measure 1M / 64K = 15 members.

There is also an important twist here: We have to measure the content of each bucket, even when it is zero. This is why I use a left join in the query above, we have to measure all the results we know, not just count the hash buckets that actually get filled (you can see how not bucketing the data would have made the test result very large).

We can now test the hypothesis against the measured result. If we see correlation, then it gives us confidence in the correctness of the hypothesis, and lets us conclude that our hash function is a good one. It also gives us the ability to compare different hash functions with each other. We can test for correlation between the predicted result and the actual result with the chi-square test. Interestingly, the tools we need for this test is even exposed in Excel as the CHISQ series of functions.

Without further ado (and insults), here are the results:

(source data available on request)

Some observations about the data:

• Most importantly: the CHECKSUM and BINARY_CHECKSUM functions are very poor hash functions. Even with 1M input and 64K large buckets, they don’t even fill the hash space. The skew is massive on non-skewed input data.
• We also see that casting the INT to NVARCHAR improves the behavior of both BINARY_CHECKSUM and CHECKSUM, but not enough to make them good hash functions.
• SHA and SHA1 seem to give us the best spread of the data, but at 1 microsecond per hash, it seems rather expensive.
• Notice that the MD series do not give a particularly nice spread. I suspect this is related to the way I divide the data with MaxInt to get from the 120bit value to a 4-byte integer
• We see that only the modulo does a distribution that gives us nearly 100% confidence (not fully though, because 1M does not divide 64K, which is reflected in the 824 value of Chi-squared). This result is expected, and my function here is engineered to “game the test”. The modulo cannot generally be used like this to spread the value all over the 32 bit integer space.
• For fun, I put in a test where I use random values all over the integer space. Even with a confidence interval of 10%, we cannot say it distributes evenly. This means that SHA functions distribute better than random on this input dataset. Your mileage will of course vary for random values

On the note of the CHECKUM series, Books Online does (under?) state:

If one of the values in the expression list changes, the checksum of the list also generally changes. However, there is a small chance that the checksum will not change. For this reason, we do not recommend using CHECKSUM to detect whether values have changed, unless your application can tolerate occasionally missing a change. Consider using HashBytes instead. When an MD5 hash algorithm is specified, the probability of HashBytes returning the same result for two different inputs is much lower than that of CHECKSUM.

Summary

In this blog I have explored properties of hash function built into SQL Server. The results indicate some general guidance that could be followed:

• For best spread over the hash space, use SHA or SHA1 algorithms with HASHBYTES function
• Be VERY careful about using CHECKSUM and BINARY_CHECKSUM as these functions do not spread a hashed integer value nicely over the full space. Depending on what you use these functions for, they may not be suitable
• It takes about 1 CPU microsecond to calculate a single call to HASHBYTES, except when hashing with MD2
• Using HASHBYTES to spread out values over the 32-bit integer space seems rather expensive, since even the best cases uses tens of thousands of CPU instructions.
• This indicates that it may be possible to implement a better hash function if you do not care about cryptographic properties of the hash

Reference:

• Grade of the Steel Category

I am happy to announce that companion volumes for performance tuners and operations people are now again, for the first time since Analysis Services 2000, a reality.

As a developer, you can learn about building, tuning and troubleshooting Analysis Services 2005, 2008 and 2008R2 cubes(yes, the guides cover all three editions) in the Performance Guide.

If you are the DBA or operations team, you can read about running such cubes in production in the Operations Guide.

And finally, if you are a consultant, expert developer or cube DBA, you can learn how to build the meanest and largest cubes from the 5 day Analysis Services Maestro Course.

These three artifacts, which I am proud to have contributed to (and for the guides, leading the effort on), are a big milestone in my Analysis Services career. They represent a large amount of knowledge transfer from Microsoft to the field. The publication of the companion volumes also marks my transition into some new and exiting projects at least for the near future. I will busy be digging into more “grade of the steel” work,  among it ROLAP UDM testing, and I hope to blog about over at SQLCAT.com very soon.

Thanks to everyone for the incredible feedback during the writing of these guys. The Analysis Services community is very vibrant and these guides are the result of our collaboration in the field, and I am happy to give something back.

References:

• Analysis Services 2008R2 Performance Guide
• Analysis Services 2008R2 Operations Guide

Throughout history, similar ideas tend to surface at about the same time. Last week, at SQLBits 9, I did some “on stage” tuning of the Paul Randal INSERT challenge.

It turns out that at almost the same time, a lab run was being done that demonstrated, on a real world workload, a technique similar to the one I ended up using. You can find it at this excellent blog: Rick’s SQL Server Blog.

Now, to remind you of the Paul Randal challenge, it consists of doing as many INSERT statements as possible into a table of this format (the test does 160M inserts total)

CREATE TABLE MyBigTable (
    c1 UNIQUEIDENTIFIER ROWGUIDCOL DEFAULT NEWID ()
    ,c2 DATETIME DEFAULT GETDATE ()
    ,c3 CHAR (111) DEFAULT ‘a’
    ,c4 INT DEFAULT 1
    ,c5 INT DEFAULT 2
    ,c6 BIGINT DEFAULT 42);

Last week, I was able to achieve  750K rows/sec (runtime: 213 seconds) on a SuperMicro, AMD 48 Core machine with 4 Fusion-io cards with this test fully tuned. I used 48 data files for best throughput, the subject of a future blog.

Now, I suspect Paul used NEWID() for key generation, to create as much I/O as possible with page splits. If the goal is purely to optimize throughput, why not ask: Is NEWID really the best way to generate a key?

Traditional SQL Server knowledge dictates that either IDENTITY(1,1) or NEWSEQUENTIALID are better key choices: they lay out the keys sequentially, at the end of the index. But one should not easily trust such recommendations without validating them. I did the above test on the 48 core box Fusion-io lend me, and here are the results:

 

This might shock you! The “Traditional, sequential layout” of the key is more than 100 times SLOWER than completely random inserts!

The CPU are NOT at 100% during this (they were when using NEWGUID). The problem shows itself when we dig into sys.dm_os_wait_stats. (by the way, below is my new: “this is how I think about it” diagram, let me know if you like this way of doing things and I will continue to use this format for illustrating troubleshooting mind processes)

We are waiting for a PAGELATCH on the last pages of the index. This wait dominates the run. The price of page splits is minor (when you have a fast I/O system) compared to the price of coordinating the last page memory access!

As Rick writes in his blog, there are ways to work around this problem by being smart about your key generation. Basically, you want a key that is NOT sequential. Rick flips the bits around in a Denali SEQUENCE objects – a very neat idea that uses a new SQL Server Denali feature.

For my test at SQLBits (which ran on 2008R2), I used a variant of the same idea: I took the @@SPID of the users session and added an increasing value (increments controlled by the app server) to each insert made. On reflecting, using the SPID is not a good idea, you should probably generate the key something like this:

INSERT Key =

   [Unique App Server ID] * 10**9 + [Unique Number Generated at App Server]

I also tried the hash trick I have described here – to compare the techniques.

Below are the results of my test of Paul’s INSERT Challenge with different key strategies (I have taken out the sequential ones, since we have seen how poorly they perform)

The last column is a trick I will describe in more detail in a later blog. Suffice to say that I reached 1.1M INSERT/sec (runtime: 150 sec) using a key generation strategy very similar to Rick’s.

Edits:

• The UNIT in the introduction was originally wrong. it should of course be 750K not 750M as can be seen later in post.

I know some of you track Conor Cunningham’s blog, which I highly recommend. The information on query optimizers is rather sparse out there – and it is a real privilege when people who know what they are talking about share their information freely on a blog. There are just not that many of them.

Here is something to think about in that context: Even if you are only interested in SQL Server, there are good things to be learned from studying other database engines – they are after all very similar. One of the good blogs for that purpose is the Oracle optimizer blog: http://blogs.oracle.com/optimizer/. It is delightfully free from marketing, and packed with good stuff. Don’t worry about the fact that they like textual query plans over there in Oracle. You might be used to viewing plans graphically, but I am sure you agree that such minor difference in display preferences can be set aside in the name of reading good material.

For those of you who would like to hear me throw my usual tantrums at conferences about data modeling, high scale tuning and optimal use of hardware. Here is my speaking schedule for the rest of 2011.

• SQLUG Sweden (14th September, Gothenburg) – Just around the corner from my home. I will be presenting about Data Warehouse tuning there.
• SQLBits – Query across the Mersey (29th September to 1st October, Liverpool) – large SQL Server conference in Europe. I really like going there: it is local, casual atmosphere, focused on tech level 400 and the place to have great discussions over beers at the conference. Last BITS I experimented with a different presentation style – this time I will go back to the old and proven style that has served me well for BITS before. I got access to some crazy hardware, so I may reveal some nasty tricks you can play with big machines.
• PASS Summit 2011 (11-14 October, Seattle) – This event, because if its location in Seattle, makes it easy for you to meet the developers of SQL Server. There will be tons of sessions. Go catch your favourite developer on the hallways after their session.
• PASS SQLRally Nordic (8-9 November, Aronsborg, Near Stockholm) – The big Nordic PASS conference. This is a new initiative hosted by the Nordic MVP and SQL community. I look  much forward to this event.

I am unfortunately locked down fully for speaking arrangements, but hopefully, I can get much more speaking done in 2012.

This week, I welcome friend, coffee connoisseur, large cube runner and coder Pete Adshead to the blogosphere.

Pete is blogging at: http://peteadshead.wordpress.com/.  His first blog is about managing SSAS through AMO – and the “interesting” coding pattern you have to adopt.

Have fun reading