SQL Server Admin 2014 :: Indexes With More Writes Than Reads
Jul 17, 2015
I have inherited a database that is over-indexed, i.e. there are sometimes 10-20 indexes on a table. The performance is at times not great due to blocking from long running queries. I want to clean up the indexes as a starting point.
Through a query I found some time ago on the SQLCat blog I have discovered a large number of indexes in the database that have a huge disparity between reads and writes. The range of difference is sometimes almost 2 million more writes than reads. Should I just drop the indexes that have say, more than 100,000 more writes than reads and then see what the Missing Index DMVs tell me after a few days of running without those indexes?
In some cases there are a few hundred thousand reads but maybe a million writes on the index. Thus, there are a fair number of reads happening, just not in comparison to the number of writes. In some cases there are almost no reads and a million or more writes. I am obviously dropping those indexes. I just am not sure what to do about the indexes that do have a fair number of reads.
Im backing up to a network directory thats actually a mount point on a different server.My backup was slower than usual so i opened up perfmon to have a look.
When selecting the mount point from the Logical Disks section in perfmon i can see that writes/sec & write bytes/sec both show zero for a long period of time, even though the backup percent complete is increasing.Then all of a sudden the writes to the network share jump massively.
Is there some caching mechanism for backups in sql where during a backup data is only flushed to the disk periodically during backup?
I read , When sql server Database having multiple data files within single filegroup then sql server writes data in multiple proportional file algorithm where the amount of data written to a file is proportionate to the amount of free space in that file, compared to other files in the filegroup.
so if there is no filegroups created and multiple secondary files are attached in databse , is there same way data stored and writes data in multiple files by the same algorithm or any different way.
The first test writes 5000000 rows in one table. I realise this is not representative OLTP behaviour, but it worked me to start interpreting performance counters and to test several setups to be discussed with our server, storage and network administrators. This way we have been able to compare the results of different hard disks, Lun vs vmdk, 1GB vs 10GB network, AMD vs Intel, etc. This way I can also compare several SQL setups (recovery model, max memory config, ...)
The screenshot shows the results of 2 runs on the same server : Win2012R2, SQL2014, 16GB RAM.
In test 1 min/max server memory was set to 9215MB/10751MB In test 2 min/max server memory was set to 13311MB/14847MB
The script assures the number of bytes inserted in the nvarchar columns is always the same.
This explains why the number of pages and the number of MB in the table are the same at the end of the 2 tests (column 5 and 6)
Since ca 13GB has to be written, the results of test 1 show the lead time is increasing once more than 10GB has been inserted (column 8 and 9) In addition you can see at that moment
- buffer cache hit ratio is decreasing - page life expectance becomes "terrible" - free list stall/sec increases - lazy writes/sec increases - readlatency increases (write latency does not)
In test 2 (id 3 in column 1 in the screenshot) those counters are not really influenced (since the 5000000 rows can all be stored in memory).
Now what I do not understand is :
Why the number of pages read (instance level) as well as the number of bytes read and the number of reads (databaselevel) is increasing extremely during run 1.
I expected to see serious impact on write behavior, since SQL server is forced to start flushing dirty pages once memory is filled. Well actually you can see here the number of writes (not the the number of bytes written) starts to increase faster in test 1 after 4000000 rows, but there's no real impact on write latency.
Finally I want to notice
- I'm the only user on this machine - the table has a clustered index on a identity column - there are no foreign key constraints - inserts are executed using a loop, not one big transaction - to monitor progress and behaviour/impact, each 10.000 loops the counters are stored using dmv queries
So I wonder why SQL Server starts to execute so many reads in test 1.
We are in the process of moving existing clustered SQL server databases to AWS. There is one major database that has intensive reads and writes transactions. I'm wondering what is the best design to optimize the performance for both R/W since we have constant issues historically with the current environment when massive updates are happening. Reads shall have higher priority over writes.
I am looking into various options to improve latency of our application (we figured the latency is mainly because data persistence - writes and reads from DB). I am looking into In-Memory databases also. But, before making that decision (of using in memory databases), I would like to see if there is a way to configure SQL Server 2005 to get as close performance as in-memory databases?
My question: 1. Is there a way that I can configure SQL Server 2005 to use a CACHE that gets loaded as needed basis, so that future database reads/writes will happen to the cache as opposed to disk (db writes)? 2. Is SQL Server 2005 recoverable in such configurations? 3. Are there any ideas/resources where I can get more details? (Such as sample configurations with bench mark numbers, rpevious experiences..etc)
I've been trying to get a definitive answer to this question but alas I have conflicting and patchy answers so far from other sources. I have an index that, lets say, requires 10GB of data space to rebuild..This index resides on a filegroup that spans 2 files on two seperate drives (i.e. a mdf and ndf)
When I rebuild this index how will each of these datafiles grow as the rebuild proceeds to completion? Lets for the time being remove the caveats of any other activity hitting the example index/database in question.My tests seem to show that only the mdf will grows (or the file with the lowest id in the that filegroup) provided there is enough space available in that particular file to complete the operation. The secondary ndf dat file doesnt grow at all if the mdf has enough space.
Is expected behavior? i.e. the index will be rebuilt in a contiguous manner relative to the files contained with the filegroup i.e. fileid 1 will grow till limit reached then next fileid grows etc?
How can You find the reads and writes per second of your hard drives in sql. I am reading my SQL book and it says that your average disk should have 125 or less i/o's. And it gave the forumal but as mentioned I don't know how to find the reads and writes.
Is there a way to get a total count of all SELECT, UPDATE, DELETE and INSERT statements to a SQL Server 6.5 database during a 12 hour period? I'm thinking maybe someone knows of a software that reads the log or monitors the server... I've been looking at the performance monitor and, although it has good information, it doesn't capture DML's.
Lets say user A accesses a record and is making an update to a column... next user B accesses the same record and makes an update to the same column and saves the data... how can user A check to see if an update has been made to prevent overwriting the data..
Is there a query statement that user A can write to check for this?
I understand locking can be used to prevent this but is there an alternative to locking.
I have an application that is insertting thousands of records houlry. The server's hard drives are staying maxxed out. My boss says there is an index problem. I say it is a drive subsystem issue.
Any help would be appreciated to understand this performance problem.
I want to set up a database role so that users can use sp_readerrorlog through SSMS. It does a check on membership in the securityadmin role.
I have tested it and can see you can grant execute on xp_readerrorlog but the SSMS GUI uses sp_readerrorlog.
I thought I could create a user/certificate and add the signature to sp_readerrorlog but it's not permitted (likely because it's not a normal database object).
So the other solution is to add the users to the securityadmin role but then explicitly deny alter any login (best done with a custom server role in 2012+ but otherwise just manually in 2008). I tested this out and it works, I'm not able to alter any logins or increase my own permissions, I also did a check of what's reported from fn_my_permissions(null, null) and it shows minimal permissions like I'd expect.
I often use profiler as one tool to identify bad plans. The reads column gives me a good indication of excessive IO to dig into and correct if necessary. I often use it with Showplan so I can see what a query does, replicate it and fix it.
However I have just lost some faith in it. I am looking at a poorly performing query joining five tables. A parallel plan has been generated and one table is being scanned (in parallel) due to a missing index. This table had in excess of 4 million rows in it. The rest hitd indexes well. However the entire query generates ONLY 12 READS.
Once corrected, a single processor plan is used. This looks really efficient and uses 120 reads. That looks the right figure to me.
Clearly 12 reads is wrong. Does the profiler only display one thread of a parallel plan perhaps? Or something else?
I'm just beginning to experiment with memory optimised tables.
I have two sets of near identical tables - one set normal, the other set memory optimised with DURABILITY=SCHEMA_ONLY - and am running test queries against these. When I say that the two sets are "near identical", I mean that they are the same except for the primary keys: for the normal tables these are defined as PRIMARY KEY CLUSTERED whereas for the memory-optimed ones they are defined as PRIMARY KEY NONCLUSTERED HASH WITH (BUCKET_COUNT=nnnn) as per the requirements for such tables.
I then run a pair of test queries, again identical but one referencing the normal tables and the other referencing the memory optimised ones.
(The query uses an inner join on three tables with row counts of approx 3m rows, 100000 rows and 5000 rows.)
The query against the normal tables runs noticeably faster than that against the memory optimised ones. To try to find out why, I examined the execution plans. the plan for the memory optimised query suggests that I have a missing index: but of course I can't create this againsty a memory optimised table. Is this a bug or am I missing something? Why the performance between the two should be so different?
If I install an instance with Windows Only authentication, and then change it to Mixed Mode, if I enable the sa login, the password has already been set. What is the default? If it's generated, how secure is it? Is the password generated? What algorithm is used for that?
My sql databases in SQL Server 2014 has the status "suspend" as I saw in SQL Management Studio. I can't restore to serviceable condition sql databases through standard procedures. I need to restore .mdf file.
I am using a monitoring system where I can monitor a numeric SQL result assuming the result is one field and one row.I would like to do this to say monitor the free available space or percentage on say the Master database. DBCC SQLPERF gives me a few columns and results for all databases on the server.
In our environment applications are using a DNS name which points to the physical server ip address. Now we are planning to move to 2014. We are planning to have servers in different subnets so we will be having two ip adresses for listener. How we can point the DNS to the listener ips? If failover happens can the DNS point to the exact ip address of the listener where it's primary node?
"Process 0:0:0 (0x1e10) Worker 0x00000006B6D341A0 appears to be non-yielding on Scheduler 13. Thread creation time: 12906028806348. Approx Thread CPU Used: kernel 0 ms, user 0 ms. Process Utilization 13%. System Idle 84%. Interval: 70189 ms."
Is it better to run the profiler or performan counter?
What are the filters we have to select in the profiler to monitor the Sql server
I have a SQL server box running 2014 reporting services. I have another server running IIS v8.
I would like to be able to connect to the IIS site and be given the SSRS report browser.
So externally if I browse to [URL], I am presented with the report server interface, the same as if I browse to http://xxx.xxx.xxx.xxx/reports internally.
What is the best approach for a read only copy of a database that is ~ 1TB. The primary database is fed nightly with an ETL process. We are currently trying to duplicate the ETL to read only server but that process is not going well. So we are looking at other options to let SQL make the copy.
The primary database is on a Win12R2 with SQL 12 or 14, a 2 node A/P failover cluster.
The read only copy will be on a Win12R2 with SQL 12 or 14. It is not a requirement to fail over to the read only copy if the primary should go down.
What would best the approach to accomplish the end result?
