We've got SQL Server 2016 (v13.0.4206.0), by default there is no restrictions for parallelism - any count SQL wants. And it didn't lead any problems... Till now.
For another feature there were written query that unexpectedly raised timeout exception in our application. I was deeply surprised when it was successfully executed with setting up maximum threads per query to 1. Yes, 6 seconds for query is not so good, even accounting to most of time was spent for fetching, but it's far away from 3 minutes timeout!
By the way, executing this query with SQL Server Management Studio works all the time despite of parallelism settings. It seems that something wrong with connection to database, but all other queries works fine, even which much harder then that one.
Our application is built on ASP.NET Core 3.0 (don't know if it matters), database connection is made using System.Data.SqlClient v4.8.0. All I could determine is that there are so much tasks created for this query:
I've tried to watch for execution in sys.dm_os_waiting_tasks (thanks google). I'm not sure I got it right, but it seems that tasks with context_id 0-8 is blocked with those who have context_id 9-16 and vise versa. Obvious example of deadlock, isn't it? But how can SQL Server manage threads to make it without my "help"? Or what am I doing wrong?
Just in case some inappropriate answers:
I won't turn parallelism off (set maximum threads per query to 1) as solution because of some heavy queries in our application;
I don't want to raise Cost Threshold for Parallelism setting because I'm afraid of same problem with another query (guess, a heavier one). So I just want to determine real cause;
Optimizing the query isn't considered (anymore), as according to actual execution plan I can't make it faster - there are enough indexes for it. But I'm ready to rethink after some really weighty arguments.
So, my question is: why does parallelism that I didn't ask for spoil the query execution? And how can I avoid that?
It's true sometimes the engine chooses to use parallel execution (or not to use) which leads to worse performance.
You do not want to control the server option and the cost as you are not sure how this will reflect to other queries, which is understandable.
If you are sure, your query will be execute better without being handle in parallel, you can specify the option just for it using query hints - MAXDOP like this:
SELECT ...
FROM ...
OPTION (MAXDOP 1);
It's easy and you can rollback if needed. Also, you are not affecting other queries.
You are saying that:
Optimizing query isn't considered (anymore), as according to actual execution plan...
The execution plan is sometimes misleading. As a start - you can save your execution plan and open it with SentryOne Plan Explorer - it's free and can give you a better look of what's going on.
Also, if a query is execute for either 3 seconds or 6 minutes, there must be something wrong with it or may be the activity of your database. If it is executed fast in the SSMS always, maybe the engine is using the correct cache plan. I thing it's better to share the query itself and to attach the two plans (serial and parallel) and spend more time tuning it.
Related
I have a few "inefficient" queries that I am trying to debug on Azure SQL (v12). The problem I have is that after the query executes for the first time (albeit, many seconds) Azure appears to cache the query / execution plan. I have done some research and several people have suggested adding and removing a column will clear the cache but this doesn't seem to work. If I leave the server alone for a few hours / overnight and re-run the query it takes its usual time to execute but once again the cache is in place - this makes it very hard to optimise my query. Does anyone know how to force Azure SQL to not cache my queries / execution plans?
ALTER DATABASE SCOPED CONFIGURATION CLEAR PROCEDURE_CACHE is designed to help wit this problem.
https://learn.microsoft.com/en-us/sql/t-sql/statements/alter-database-scoped-configuration-transact-sql?view=sql-server-2017
This is closest to the DBCC FREEPROCCACHE you have in SQL Server but is scoped to a database instead of the server instance. This does not prevent caching of query plans - it just invalidates the current cache entries.
Please note that the query store is there to help you in SQL Azure (on-by-default). It stores a history of plan choices and plan performance (per-plan). So, if you have a prior plan that performs better available in the history of your application, you can force it using SSMS if you'd prefer to have the query optimizer pick this plan each time your query compiles. One common reason for what you are seeing is parameter-sensitivity in the plan choice where the optimizer will use the passed parameter value to try to generate the query plan, assuming it is representing a common pattern when you run that query. If that value is actually not close to a common value (in terms of how frequent it is in the table), then you can sometimes compile and cache a plan that is not better on average for your application.
Query store has an overview here:
https://learn.microsoft.com/en-us/sql/relational-databases/performance/monitoring-performance-by-using-the-query-store?view=sql-server-2017
Note that SQL Azure also has an automated mechanism to try forcing prior plans if it notices a performance regression. It is somewhat conservative, however, so it may not kick in for every single regression until it sees an obvious pattern over time. So, while you can force things in SSMS, you can also potentially just wait (assuming this is the issue you were seeing)
I am using SQL Server 2008 R2.
The process is actually like this:
First, about 2 million records are pulled from a remote server,
then a join is done locally,
the final result is thousands of records.
The time cost varies from less one 1 min to 30 mins.
And after I experienced the 30 mins delay, it seems the following time costs are all only around 3 mins.
It is the same data, same SP.
What could cause this drastic difference?
Update
I delet the SP, re-start the SQL server service, and re-creat the SP. The execution took only 50 seconds!
What's wrong?
The behaviour you describe seems extreme - but (if you exclude the client), there are 3 logical places to look.
The first is the query execution on the database server. It's worth using the Query Analyzer tool to see if it's using any indices - by far the most common reason for variable performance of database queries is that the query is not using (the right) indices, and that therefore the impact of the query cache plays a big part. SQL Server will cache a lot of data, and the first run of your proc populates that cache; the second run is faster because it hits the cache. After a while, the cache goes stale, and running the proc slows down again.
The second possibility is that the database server is wobbly - it may just not be powerful enough to do all the work it's supposed to do. In that case, one moment you get lucky, have all the server resources to yourself; the next, someone else is running a query and yours slows down. That would make all queries slow, not just this one - so it doesn't sound likely.
Third possibility is networking weirdness - as Phil says, "thousands of records" is nothing too scary, but if they're big, and your network is saturated with pictures of kittens, it might have an impact. Again, that would manifest in general network slowness, and is unlikely to explain a delay of 30 minutes...
Fourth, is anything going on at the same time?
Fifth, does your SP use dynamically generated SQL statements? This would cause the SP not to become pre-compiled. If possible seperate such statements into child SPs.
I apologize in advance for not having all of the specifics available, but the machine is building an index probably for a good while still and is almost completely unresponsive.
I've got a table on SQL Server 2005 with a good number of columns, maybe 20, but a mammoth number of rows (tens, more likely hundreds of millions). In order to simplify the amount of JPA work I'd need to do to access it, I created a view that contained the bits I was interested in. The view was created as:
SELECT bigtable.ID, bigtable.external_identification, mediumtable.hostname,
CONVERT(VARCHAR, bigtable.datefield, 121) AS datefield
FROM schema.bigtable JOIN schema.mediumtable ON bigtable.joinID = mediumtable.ID;
When I want to select from the view, I do:
SELECT * FROM vwTable WHERE external_identification = 'some string';
This works just fine in SQL Management Studio. The external_identification column has a non-unique, non-clustered index in bigtable. This also worked just fine on our remotely executing Java program in our test environment. Now that we're a day or two away from production, the code has been changed a bit (although the fundamental JPA NamedQuery is still straightforward), but we have a new SQLServer installation on new hardware; the test version was on a 32-bit single core machine, the new hardware is 64-bit multi-core.
Whenever I try to run the code that uses this view on the new hardware, it either hangs indefinitely on the first call of this query or times out if I have a timeout specified. After doing some digging, something like:
SELECT status, command, wait_type, last_wait_type FROM sys.dm_exec_requests;
confirmed that the query was running, but showed it in the state:
suspended, SELECT, CXPACKET, CXPACKET
for as long as I cared to wait for it. Whenever I ran the exact same query from within the Management Studio, it completed immediately. So I did some research, and found out this is due to waiting on some kind of concurrent operation to start/finish. In an attempt to circumvent that, I set the server-wide MAXDOP to 1 (disabled concurrency). After that, the query still hangs, but the sys.dm_exec_requests would show:
suspended, SELECT, PAGEIOLATCH_SH, PAGEIOLATCH_SH
This indicates that it's some sort of HD/scanning issue. While certainly the machine is less responsive than I'd expect for newer hardware, I wouldn't expect this query (even over the view) to require much scanning, since the column I'm searching by is indexed in the underlying table and it works if I run it locally. But just because I'm out of ideas and under the gun, I'm adding indexes to the view; first I have to add the unique clustered index (over ID) before I can attempt to add the non-unique non-clustered index over external_identification.
I'm the only one using this database; when I select from sys.dm_exec_requests the only two results are the query I'm actively inspecting and the select from sys.dm_exec_requests query. So it's not like it's under legitimately heavy, or even at all concurrent, load.
But I suspect I'm grasping at straws. I'm no DBA, and every time I have to interact with SQL Server outside of querying it it baffles my intuitions. Does anyone have any ideas why a query executed remotely would immediately go into a suspended state while the same query locally would execute immediately?
Wow, this one caught me straight out of left field. It turns out that by default, the MSSQL JDBC driver sends its String datatypes as Unicode, which the table/view might not be prepared to handle specifically. In our case, the columns and indexes were not, so MSSQL would perform a full table scan for each lookup.
In our test environment, the table was small enough that this didn't matter, so I was tricked into thinking it worked fine. In retrospect, I'm glad it didn't -- I can't stand it when computers give the illusion of inconsistency.
When I added this little parameter to the end of my JDBC connection string:
jdbc:sqlserver://[IP]:1433;databaseName=[db];sendStringParametersAsUnicode=false
things immediately and magically started working. Sorry for the slightly misleading question (I barely even mentioned JPA), but I had no idea what the cause was and really did believe it was something SQL Server side. Task Manager didn't report heavy CPU/Memory usage while the query was suspended, so I just thought it was idling even though it was really under heavy disk usage.
More info about MSSQL JDBC and Unicode can be found where I stumbled across the solution, at http://server.pramati.com/blog/2010/06/02/perfissues-jdbcdrivers-mssqlserver/ . Thanks, Ed, for that detailed shot in the dark -- it may not have been the problem, but I certainly learned a lot (and fast!) about MSSQL's gritty parts!
It is likely that the query run in SSMS and by your application are using different query plans - from the wait types you're seeing in dm_exec_requests it sounds like the plan created for the application is doing a table scan where the plan for SSMS is using an index seek.
This is possible because the SSMS and application database connections likely use different connection options, some of which are used as a key to the database plan cache.
You can find out which options your application is using by running a default SQL server profiler trace against the server; the first command after the connection is created will be a number of SET... options:
SET DATEFORMAT dmy
SET ANSI_NULLS ON
...
I suspect this list will be different between your application and your SSMS connection - a common candidate is SET ARITHABORT {ON|OFF}, since that forms part of the key of the cached plan.
If you run the SET... commands in an SSMS window before executing the query, the same (bad) plan as is being used by the application should then be picked up.
Assuming this demonstrates the problem, the next step is to work out how to prevent the bad plan getting into cache. It's difficult to give generic instructions about how to do this, since there are a few possible causes.
It's a bit of a scattergun approach (there are other more targetted ways to attempt to resolve this problem but they require more detailed understanding of the issue that I have now), but one thing to try is to add OPTION (RECOMPILE) to the end of your query - this forces a new plan to be generated for every execution, and should prevent the bad plan being reused:
SELECT * FROM vwTable WHERE external_identification = 'some string' OPTION (RECOMPILE);
Assuming you can replicate the bad performance is SSMS using the steps above, you should be able to test this there.
Beware that this can have negative performance consequences if the query is executed very frequently (since each recompilation requires CPU) - this depends on the workload of your application and will need testing.
A couple of other thoughts:
Check the schemas between the test and production systems; this might be as simple as a missing index from one of the tables in the production database, although given that SSMS queries perform OK this is unlikely.
You should re-enable parallelism by taking the server-wide MAXDOP=1 off, since this will limit the performance of your system overall. The problem is almost certainly the query plan, not parallelism
You also need to beware of the consequences of adding indexes to the view - doing so effectively materialises the view, which will (given the size of the table) require a lot of storage overhead - the indexes will also need to be maintained when INSERT/UPDATE/DELETE statements take place on the base table. Indexing the view is probably unnecessary given that (from SSMS) you know it's possible for the query to perform.
I'm running into a performance issue with the current schema. So I built an equivalent schema to solve the issue.
I ran some tests on both schemas and the results are hard to understand. For the record, the data is the same.
I get the following from the Profiler when executing equivalent requests on the two schemas.
Old schema:
1,300,000 reads
5,000 CPU
4 seconds execution time
New schema:
30,000 reads
3,000 CPU
6 seconds execution time
The difference seems to be in the query plan used. The old schema has parallelism in the query plan. The new schema isn't using parallelism.
Has anyone faced similar situations (less IO/CPU but more execution time). How did you solve it?
Is there a way to force parallelism? I've played with query hints(http://msdn.microsoft.com/en-us/library/ms18171). I'm able to stop parallelism on the old schema but can't seem the query on the new schema to use parallelism.
Thanks in advance.
Louis,
Currently there is no way to force parallelism in SQL Server straight out of the box but Adam Machanic did some work to do that though.
http://whoisactive.com
Coming to your first question, yes we have seen cases like that too. Note that Parallelism is cpu bound and that's why you are seeing more cpu time but overall less execution time as you have multiple threads doing the work for you.
http://www.simple-talk.com/sql/learn-sql-server/understanding-and-using-parallelism-in-sql-server/
Make sure you have proper indexes in place and also stats are updated with full scan. In the long run it is best if Query Optimizer makes the decisions by itself but if you want to overwrite the QO plans then you may have to add lot more details. Schema, data and repro.
HTH
It's my understanding that SQL Server 2005 does some sort of result or index caching. I'm currently profiling complex select statements which take several seconds to several minutes to complete. My problem is that a second run of a query never takes more than a second to run even if I don't alter it. I'm currently using SQL Server Management Studio Express to execute the queries against a SQL Server 2005 server.
My question is, Is there any way to avoid or clear the cache that is causing my queries to execute so quickly on a second run?
There are a couple different things that could at play here, the 3 that come to mind initially (in probable-ish order) would be as follows - if you would like some help interpreting the results, follow the instructions below and paste the stat outputs in the question:
Your query/batch is taking a long time to compile an execution plan. Execution plans are determined and cached (see this post on serverfault for an overview of understanding for how long, when they are rebuilt, etc.)
To verify this, turn on statistics time output, which will provide you information on how long the engine is taking to generate a query plan. For the query/batch in question:
DBCC FREEPROCCACHE
SET STATISTICS TIME ON
Execute the batch, capture the stats output
Execute the batch again, capture the stats output
Compare the 2 stat outputs, paying particular attention to the parse/compile time differences between the 2 executions.
If this is the problem, you can take a couple of approaches to resolving the issue, including specifying a plan guide, specifying a static plan with use plan, or possibly other options like creating a scheduled job to simply compile the plan every few minutes (not as good on option on Sql 2k5 as the others).
Your query/batch is touching a lot of data - on the first execution the data may not be in the buffer pool (basically the cached pages of data the server needs) and the query is performing physical IO operations as opposed to logical IO operations (i.e. reads from disk vs. reads from cache).
To verify this, turn on statistics io output, which will provide you information on the types of IOs and how many of those the engine is performing for the batch. For the query/batch in question:
DBCC DROPCLEANBUFFERS
SET STATISTICS IO ON
Execute the batch, capture the stats output
Execute the batch again, capture the stats output
Compare the 2 stat outputs, paying particular attention to the physical/read-ahead and logical IO outpus between the 2 executions.
To resolve this, you've basically got only 1 option - optimize the query in question so it performs fewer IO operations. You could consider creating a scheduled job that runs the query every so ofter to keep the data in the buffer pool, but this wouldn't be as good an option.
Your query/batch is getting a poor execution plan and/or a poor execution plan choice for different variable values - is this a batch/query that is using a parameterized statement (i.e. you are using variables/static values in the where/join clauses?)? If so, are you seeing the difference in execution times for the same values or different values? If for the same values, the answer is likely #1 or #2 - if for different values, this is potentially your problem. If you think this is the issue after researching #1 and #2, repost with the .sqlplan, the TSQL, and the different parameter values you are using.
I've found the only realiable metrics for performance tuning come from SQL Server's Profiler application. When looking at CPU Time, and Reads in particular, you become much more sheltered from 'other influences'.
For example, the OS being busy, or multiple users being active will reduce your share of CPU and so increase the time to execute. And you may or may not get parallelism over multiple CPUs. But either way, the total CPU Time (as opposed to execution time) will stay approximately the same.
Do this:
CHECKPOINT
DBCC DROPCLEANBUFFERS