Database Monitoring

SQL Server - Performance Counter 


SQLServer:Access Methods (Forwarded Records/sec): Rows with varchar columns can experience expansion when varchar values are updated with a longer string.  In the case where the row cannot fit in the existing page, the row migrates and access to the row will traverse a pointer.  This only happens on heaps (tables without clustered indexes). Evaluate clustered index for heap tables.  In cases where clustered indexes cannot be used, drop non-clustered indexes, build a clustered index to reorg pages and rows, drop the clustered index, then recreate non-clustered indexes.
Preferred Value(< 10 per 100 Batch Requests/Sec)

SQLServer:Access Methods (Full Scans / sec):  This counter monitors the number of full scans on base tables or indexes. Values greater than 1 or 2 indicate that we are having table / Index page scans. If we see high CPU then we need to investigate this counter, otherwise if the full scans are on small tables we can ignore this counter.  A few of the main causes of high Full Scans/sec are
• Missing indexes
• Too many rows requested
Queries with missing indexes or too many rows requested will have a large number of logical reads and an increased CPU time.
Preferred Value(Index Searches/sec)/(Full Scans/sec) > 1000

SQLServer:Access Methods(Index Searches/sec):  Number of index searches. Index searches are used to start range scans, single index record fetches, and to reposition within an index. Index searches are preferable to index and table scans.  For OLTP applications, optimize for more index searches and less scans (preferably, 1 full scan for every 1000 index searches). Index and table scans are expensive I/O operations.
Preferred Value ((Index Searches/sec)/(Full Scans/sec) > 1000): 

SQLServer:Access Methods(Page Splits/sec):     Number of page splits per second that occur as the result of overflowing index pages. Interesting counter that can lead us to our table / index design. This value needs to be low as possible. If you find out that the number of page splits is high, consider increasing the fillfactor of your indexes. An increased fillfactor helps to reduce page splits because there is more room in data pages before it fills up and a page split has to occur.
Note that this counter also includes the new page allocations as well and doesn’t necessarily pose a problem.  The other place we can confirm the page splits that involve data or index rows moves are the fragmented indexes on page splits.

SQL Server:Buffer Manager(Buffer Cache hit ratio): This counter indicates how often SQL Server goes to the buffer, not the hard disk, to get data. The higher this ratio, the less often SQL Server has to go to the hard disk to fetch data, and performance overall is boosted. Unlike many of the other counters available for monitoring SQL Server, this counter averages the Buffer Cache Hit Ratio from the time the last instance of SQL Server was restarted. In other words, this counter is not a real-time measurement, but an average of all the days since SQL Server was last restarted. In OLTP applications, this ratio should exceed 90-95%. If it doesn't, then you need to add more RAM to your server to increase performance. In OLAP applications, the ratio could be much less because of the nature of how OLAP works. In any case, more RAM should increase the performance of SQL Server OLAP activity.
Preferred Value: > 90%

SQL Server:Buffer Manager(Free list stalls/sec):               Free list stalls/sec is the frequency with which requests for available database pages are suspended because no buffers are available. Free list stall rates of 3 or 4 per second indicate too little SQL memory available.
Preferred Value: > 2

SQL Server:Buffer Manager(Free pages):  Total number of pages on all free lists.
Preferred Value: > 640  

SQL Server:Buffer Manager(Lazy Writes/Sec): This counter tracks how many times a second that the Lazy Writer process is moving dirty pages from the buffer to disk in order to free up buffer space. Generally speaking, this should not be a high value, say more than 20 per second or so.  Ideally, it should be close to zero. If it is zero, this indicates that your SQL Server's buffer cache is plenty big and SQL Server doesn't have to free up dirty pages, instead waiting for this to occur during regular checkpoints. If this value is high, then a need for more memory is indicated.
Preferred Value:              < 20

SQL Server:Buffer Manager(Page Life Expectancy): This performance monitor counter tells you, on average, how long data pages are staying in the buffer. If this value gets below 300 seconds, this is a potential indication that your SQL Server could use more memory in order to boost performance.
Preferred Value: > 300

SQLServer:Buffer Manager(Page lookups/sec): Number of requests to find a page in the buffer pool. When the ratio of page lookups to batch requests is much greater than 100, this is an indication that while query plans are looking up data in the buffer pool, these plans are inefficient. Identify queries with the highest amount of logical I/O's and tune them.
Preferred Value: (Page lookups/sec) / (Batch Requests/sec) < 100

SQL Server:Buffer Manager(Page reads/sec): Number of physical database page reads issued. 80 – 90 per second is normal, anything that is above indicates indexing or memory constraint.
Preferred Value: < 90

SQL Server:Buffer Manager(Page writes/sec): Number of physical database page writes issued. 80 – 90 per second is normal, anything more we need to check the lazy writer/sec and checkpoint counters, if these counters are also relatively high then, it’s memory constraint.
Preferred Value: < 90

SQLServer:General Statistics(Logins/sec):  > 2 per second indicates that the application is not correctly using connection pooling.
Preferred Value:  < 2      

SQLServer:General Statistics(Logouts/sec): > 2 per second indicates that the application is not correctly using connection pooling.
Preferred Value:  < 2     

SQLServer:General Statistics(User Connections ): The number of users currently connected to the SQL Server.  
Note: It is recommended to review this counter along with “Batch Requests/Sec”.   A surge in “user connections” may result in a surge of “Batch Requests/Sec”.  So if there is a disparity (one going up and the other staying flat or going down), then that may be a cause for concern. With a blocking problem, for example, you might see user connections, lock waits and lock wait time all increase while batch requests/sec decreases.  

SQL Server:Latches(Latch Waits/sec):  This is the number of latch requests that could not be granted immediately. In other words, these are the amount of latches, in a one second period that had to wait.
Preferred Value:  (Total Latch Wait Time) / (Latch Waits/Sec) < 10       
       
SQL Server:Latches( Total Latch Wait Time (ms)): This is the total latch wait time (in milliseconds) for latch requests in the last second
Preferred Value:              (Total Latch Wait Time) / (Latch Waits/Sec) < 10

SQL Server:Locks  (Lock Wait Time (ms)):  Total wait time (milliseconds) for locks in the last second.
 Note: For “Lock Wait Time” it is recommended to look beyond the Avg value.  Look for any peaks that are close (or exceeds) to a wait of 60 sec.   Though this counter counts how many total milliseconds SQL Server is  waiting on locks during the last second, but the counter actually records  at the end of locking event.  So most probably the peaks represent one huge locking event.  If those events exceeds more than 60seconds then they may have extended blocking and could be an issue. In such cases, thoroughly analyze the blocking script output. Some applications are written for timing out after 60 seconds and that’s not acceptable response for those applications.

SQL Server:Locks  (Lock Waits/sec): This counter reports how many times users waited to acquire a lock over the past second.  Note that while you are actually waiting on the lock that this is not reflected in this counter—it gets incremented only when you “wake up” after waiting on the lock. If this value is nonzero then it is an indication that there is at least some level of blocking occurring.  If you combine this with the Lock Wait Time counter, you can get some idea of how long the blocking lasted.  A zero value for this counter can definitively prove out blocking as a potential cause; a nonzero value will require looking at other information to determine whether it is significant. 
Preferred Value: 0

SQL Server:Locks  (Number of Deadlocks/sec):  The number of lock requests that resulted in a deadlock.
Preferred Value: < 1      

SQLServer:Memory Manager(Total Server Memory(KB)):  The Total Server Memory is the current amount of memory that SQL Server is using.  If this counter is still growing the server has not yet reached its steady-state, and it is still trying to populate the cache and get pages loaded into memory.  Performance will likely be somewhat slower during this time since more disk I/O is required at this stage.  This behavior is normal.  Eventually Total Server Memory should approximate Target Server Memory.

SQLServer:SQL Statistics (Batch Requests/Sec): This counter measures the number of batch requests that SQL Server receives per second, and generally follows in step to how busy your server's CPUs are. Generally speaking, over 1000 batch requests per second indicates a very busy SQL Server, and could mean that if you are not already experiencing a CPU bottleneck, that you may very well soon. Of course, this is a relative number, and the bigger your hardware, the more batch requests per second SQL Server can handle. From a network bottleneck approach, a typical 100Mbs network card is only able to handle about 3000 batch requests per second. If you have a server that is this busy, you may need to have two or more network cards, or go to a 1Gbs network card.  
Note: Sometimes low batch requests/sec can be misleading.  If there were a SQL statements/sec counter, this would be a more accurate measure of the amount of SQL Server activity.  For example, an application may call only a few stored procedures yet each stored procedure does lot of work.  In that case, we will see a low number for batch requests/sec but each stored procedure (one batch) will execute many SQL statements that drive CPU and other resources.  As a result, many counter thresholds based on the number of batch requests/sec will seem to identify issues because the batch requests on such a server are unusually low for the level of activity on the server.   
We cannot conclude that a SQL Server is not active simply by looking at only batch requests/sec.  Rather, you have to do more investigation before deciding there is no load on the server.  If the average number of batch requests/sec is below 5 and other counters (such as SQL Server processor utilization) confirm the absence of significant activity, then there is not enough of a load to make any recommendations or identify issues regarding scalability.

SQLServer:SQL Statistics (SQL Compilations/sec):            The number of times per second that SQL Server compilations have occurred. This value needs to be as low as possible. If you see a high value such as over 100, then it’s an indication that there are lots of adhoc queries that are running, might cause CPU usage, solution is to re-write these adhoc as stored procedure or use sp_executeSQL.
Preferred Value: < 10% of the number of Batch Requests/Sec

SQLServer:SQL Statistics (SQL Re-Compilations/sec) : This needs to be nil in our system as much as possible. A recompile can cause deadlocks and compile locks that are not compatible with any locking type.
Preferred Value: < 10% of the number of SQL Compilations/sec

No comments:

Post a Comment