28

I'm trying to come up with a sane way to understand if the max server memory (mb) setting is appropriate (either should be lower, or higher, or stay the way it is). I am aware that max server memory (mb) should always be low enough to leave room for the operating system itself, etc.

The environment I'm looking at has several hundred servers; I need a reliable formula I can use to determine if the present size of the buffer pool is appropriate since RAM is costed per GB allocated to each server. The entire environment is virtualized, and "physical" RAM allocated to a VM can easily be changed up or down.

I have a particular SQL Server Instance I'm looking at now with a PLE of 1,100,052 seconds, which equates to 12.7 days (the amount of time the server has been up). The server has a max server memory setting of 2560MB (2.5GB), of which only 1380MB (1.3GB) is actually committed.

I've read several items including one by Jonathan Keheyias (post) and another by Paul Randal (post), and several others. Jonathan advocates monitoring for a PLE below 300 per 4GB of buffer pool as being too low. For the SQL Server Instance above, 300 * (2.5 / 4) = 187 results in a really really low target PLE below 300. This instance has 290GB of SQL Server data (not including log files), and is only used for integration testing. Assuming the last 12 days is representative of typical usage for this server, I'd say the max server memory (mb) setting could be lowered.

At the other end of the scale, I have another integration-test server with a PLE of 294, that has a max server memory (mb) setting of only 1GB. This server has only 224MB of SQL Server data not including logs, and is running some BizFlow databases. This server might benefit from a higher max server memory (mb) setting.

I'm thinking a good starting place for targets that might be assigned too much memory could include looking at:

SELECT 
    RamMB = physical_memory_in_bytes / 1048576
    , BufferPoolCommittedMB = bpool_committed * 8192E0 / 1048576
    , BufferPoolCommitTargetMB = bpool_commit_target * 8192E0 / 1048576
    , PercentOfDesiredSizeMB = CONVERT(INT,(CONVERT(DECIMAL(18,2),bpool_committed) 
                            / bpool_commit_target) * 100)
FROM sys.dm_os_sys_info;

If BufferPoolCommitTargetMB / BufferPoolCommittedMB is greater than 1, the server is not using the entire buffer pool. If the machine in question also has a PLE of greater than "x" then it might be a good candidate for a decrease in max server memory (mb).

Since the Buffer Manager:Lazy writes/sec performance counter tracks the number of times SQLOS has written pages out to disk between checkpoints due to memory pressure, this might be another good thing to look at.

DECLARE @WaitTime DATETIME;
SET @WaitTime = '00:00:15';
DECLARE @NumSeconds INT;
SET @NumSeconds = DATEDIFF(SECOND, 0, @WaitTime);
DECLARE @LazyWrites1 BIGINT;
DECLARE @LazyWrites2 BIGINT;

SELECT @LazyWrites1 = cntr_value 
FROM sys.dm_os_performance_counters dopc
WHERE (
        dopc.counter_name LIKE 'Lazy writes/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    )
    AND dopc.object_name = 'MSSQL$' + CONVERT(VARCHAR(255),
               SERVERPROPERTY('InstanceName')) + ':Buffer Manager';

WAITFOR DELAY @WaitTime;

SELECT @LazyWrites2 = cntr_value 
FROM sys.dm_os_performance_counters dopc
WHERE (
        dopc.counter_name LIKE 'Lazy writes/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    )
    AND dopc.object_name = 'MSSQL$' + CONVERT(VARCHAR(255),
               SERVERPROPERTY('InstanceName')) + ':Buffer Manager';

SELECT LazyWritesPerSecond = (@LazyWrites2 - @LazyWrites1) / @NumSeconds;

The above code assumes the server is under load during the 15 seconds it takes to run, otherwise it will report 0; which might be a misleading false-negative.

Should I also look at PAGELATCHIO_* wait stats or some other wait type as an indicator of memory pressure, or lack thereof?

My question is, how can I reliably determine a "good" target value for PLE and max server memory (mb)?

11

As you already know there is No general formula to calculate max server memory you can do some quick maths and reach to a value but still you would need help of Perfmon counters at last to monitor memory usage and change accordingly. I know below general formula and I use it as well. I learned this formula from This Link

For SQL Server 2005 to 2008 R2

Please note from SQL Server 2005 to 2008 R2 max server memory only controls buffer pool. So max server memory configuration is bit tedious here and involves few calculations

  1. Leave 2 G memory straight away for Windows OS.

  2. Of course system would have antivirus running. Please leave 1.5G for Antivirus. Please note Mcafee and SQL Server do not go hand in hand so make sure you leave enough for it. You can also check perfmon counter Perfmon Process-> Private bytes and Working Set to monitor memory usage by AV and other small applications running on SQL Server box

enter image description here

  1. Consider the memory requirements of the drivers/firmwares.You have to derive it based on memory requirements by drivers installed on the system. RAMMAP tool can help

  2. Consider the NonbPool (aka MTL or MTR) memory requirements by SQL Server.

    select  sum(multi_pages_kb)/1024 as multi_pages_mb from  sys.dm_os_memory_clerks
    

    + Max worker threads * 2MB

    + Memory for direct Windows allocations approximately 0 to 300 MB in most of the cases but you may have to increase it if there are many 3 party components loaded in SQL Server process (Including linked server dll’s, 3rd party backup dll’s etc.)

    + If you are using CLR extensively add some additional memory for CLR.

  3. Consider the memory requirement by jobs (Including replication agents, Log shipping etc. ) and packages that will run on the server. It can very from MB's to GB's according to number of jobs running. For medium sized server you can take it as 250 MB

  4. Make sure there is good enough free space for operating system.

    Approximately (100 MB for each GB till 4G) + (50 MB for each additional GB till 12GB) + (25 MB for each additional GB till your RAM size)

  5. Other memory requirements.

    If you have any other memory requirement specific to your environment.

    Max server memory= Total physical memory – (1+2+3+4+5+6+7)

    I have not included memory configuration for SSIS.SSRS,SSAS you would also need to subtract memory required by these services from total physical server memory.

    After you have configured above you need to monitor following counters

  • SQLServer:Buffer Manager--Page Life Expectancy(PLE):

  • SQLServer:Buffer Manager--CheckpointPages/sec:

  • SQLServer:Memory Manager--Memory Grants Pending:

  • SQLServer:memory Manager--Target Server Memory:

  • SQLServer:memory Manager--Total Server memory

For SQL Server 2012/2014.

From SQL Server 2012 onwards setting up max server memory has become easy. Because now max server memory almost accounts for all memory consumption. Max server memory controls SQL Server memory allocation, including the buffer pool, compile memory, all caches, qe memory grants, lock manager memory, and CLR memory (basically any “clerk” as found in dm_os_memory_clerks). Memory for thread stacks, heaps, linked server providers other than SQL Server, or any memory allocated by a “non SQL Server” DLL is not controlled by max server memory.

You can allocate 75-80% to SQL Server and then use perfmon counters to monitor memory usage. In SQL Server 2012 few perfmon counters have been deprecated. Buffer manager counter is deprecated you must use memory manager counter

  • SQL Server: Memory Manager-- Target Server Memory (KB)

  • SQL Server: Memory Manager--Total Server Memory (KB)

  • SQL Server: Memory Manager- Free Memory (KB)

  • SQL Server: Memory Manager--Database Cache Memory (KB)

On value of PLE, I have used formula by Joanthan and fortunately it has worked for me.

6

The challenge here is that numbers don't take into account end user experience.

Great example: I have database server used to track every web site that the company employees visit. I don't care if it can't keep up with inserts during peak loads because the front end app batches the inserts off periodically, and slow inserts don't cause a problem for users. Users are still able to surf the web without being held up by slow inserts.

At SELECT time, the HR department just fires off reports when asked for suspicious browsing history for a given employee, but they don't care how long the reports take - they just open the report and go off to do other things.

Performance needs to start with a question: are the users happy with performance? If so, leave the system where it is.

  • Even if you're using more memory than you need to? – James Anderson Dec 11 '14 at 12:18
  • 2
    James - generally speaking, I don't want to make changes that cause users to complain. If you do want to do that, you could gradually step down the amount of memory for each server until the users start to complain, but when I'm already overworked, I generally don't have time to take those steps. I have to focus on the tasks that will make unhappy users happy - rather than trying to make happy users unhappy. ;-) – Brent Ozar Dec 11 '14 at 15:22
  • 2
    Good points, Brent. I was asked to see if some servers are over-provisioned because we pay for memory per GB per year. Many of the instances I'm looking at have what I consider to be a very small amount of RAM in max server memory (mb), and as such I'm pretty reluctant to down-size them. However, some other instances have 1,000,000 + PLE, and as such are pretty obvious potential candidates for a drop in RAM. Clearly, lowering RAM will cause an increase in IOps, and I'm not sure what the cost of that will be. – Max Vernon Dec 11 '14 at 22:17
  • Also, looking at PLE compared to the max server memory setting is kind of a chicken-and-egg thing; the lower the max server memory setting, the lower the minimum "acceptable" PLE would be, so I could get stuck in an ever-lowering spiral. I'm sure, as you mention, the user performance will at some point be affected. – Max Vernon Dec 11 '14 at 22:19
  • PLE counters is one which you should avoid looking from 2012 onwards or when you have NUMA system where each node behaves as own small memory allocator. If you want you should look for PLE for each NUMA node not complete you might get incorrect value – Shanky Dec 13 '14 at 23:27
3

The current T-SQL I'm using to evaluate PLE vs max server memory is:

/*
    Purpose:            Returns a resultset describing various server level stats including PLE
                        Max and Min Server Memory, etc.
    By:                 Max Vernon
    Date:               2014-12-01
*/
SET NOCOUNT ON;

/*
    wait stats for PAGELATCH_IO
*/
DECLARE @Debug BIT;
SET @Debug = 0;
DECLARE @HTMLOutput BIT;
SET @HTMLOutput = 1;
DECLARE @WaitTime DATETIME;
SET @WaitTime = '00:00:15';
DECLARE @NumSeconds INT;
SET @NumSeconds = DATEDIFF(SECOND, 0, @WaitTime);
DECLARE @InstanceName NVARCHAR(255);
SET @InstanceName = CONVERT(NVARCHAR(255), SERVERPROPERTY('InstanceName'));
DECLARE @Version NVARCHAR(255);
DECLARE @VersionINT INT;
SET @Version = CONVERT(NVARCHAR(255),SERVERPROPERTY('ProductVersion'));
SET @VersionINT = CONVERT(INT, SUBSTRING(@Version,1 ,CHARINDEX('.',@Version)-1));
DECLARE @cmd NVARCHAR(MAX);
SET @cmd = '';
DECLARE @TaskCount INT;
DECLARE @TasksPerSecondAvg INT;
DECLARE @AvgWaitTimeInMSPerTask DECIMAL(10,2);
DECLARE @AvgWaitTimeInMSPerSecond DECIMAL(10,2);
DECLARE @TotalWaitTimeInMSOverall DECIMAL(10,2);
DECLARE @LazyWrites1 BIGINT;
DECLARE @LazyWrites2 BIGINT;
DECLARE @FreeListStallsSec1 BIGINT;
DECLARE @FreeListStallsSec2 BIGINT;
DECLARE @BatchReq1 BIGINT;
DECLARE @BatchReq2 BIGINT;
DECLARE @ws TABLE
(
    RunNum INT
    , wait_type SYSNAME
    , waiting_tasks_count BIGINT
    , wait_time_ms BIGINT
    , max_wait_time_ms BIGINT
    , signal_wait_time_ms BIGINT
);
INSERT INTO @ws
SELECT 1, dows.*
FROM sys.dm_os_wait_stats dows
WHERE dows.wait_type LIKE 'PAGEIOLATCH_%'
ORDER BY dows.waiting_tasks_count DESC;

SELECT @LazyWrites1 = cntr_value 
FROM sys.dm_os_performance_counters dopc
WHERE (
        dopc.counter_name LIKE N'Lazy writes/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    )
    AND dopc.object_name = N'MSSQL$' + @InstanceName + N':Buffer Manager';

SELECT @FreeListStallsSec1 = cntr_value 
FROM sys.dm_os_performance_counters dopc
WHERE (
        dopc.counter_name LIKE N'Free list stalls/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    )
    AND dopc.object_name = N'MSSQL$' + @InstanceName + N':Buffer Manager';

SELECT @BatchReq1 = cntr_value
FROM sys.dm_os_performance_counters dopc
WHERE dopc.counter_name LIKE N'Batch Requests/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    AND dopc.object_name = N'MSSQL$' + @InstanceName + N':SQL Statistics';

WAITFOR DELAY @WaitTime;

INSERT INTO @ws
SELECT 2, dows.*
FROM sys.dm_os_wait_stats dows
WHERE dows.wait_type LIKE N'PAGEIOLATCH_%'
ORDER BY dows.waiting_tasks_count DESC;

SELECT @LazyWrites2 = cntr_value 
FROM sys.dm_os_performance_counters dopc
WHERE (
        dopc.counter_name LIKE N'Lazy writes/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    )
    AND dopc.object_name = N'MSSQL$' + @InstanceName + N':Buffer Manager';

SELECT @FreeListStallsSec2 = cntr_value 
FROM sys.dm_os_performance_counters dopc
WHERE (
        dopc.counter_name LIKE N'Free list stalls/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    )
    AND dopc.object_name = N'MSSQL$' + @InstanceName + N':Buffer Manager';

SELECT @TaskCount = SUM(w2.waiting_tasks_count - w1.waiting_tasks_count)
    , @TasksPerSecondAvg = CONVERT(DECIMAL(10,2), (SUM(w2.waiting_tasks_count) - SUM(w1.waiting_tasks_count))) / @NumSeconds
    , @AvgWaitTimeInMSPerTask = CONVERT(DECIMAL(10,2),(SUM(w2.wait_time_ms) - SUM(w1.wait_time_ms))) / CONVERT(DECIMAL(10,2),(SUM(w2.waiting_tasks_count) - SUM(w1.waiting_tasks_count)))
    , @AvgWaitTimeInMSPerSecond = (CONVERT(DECIMAL(10,2), (SUM(w2.waiting_tasks_count) - SUM(w1.waiting_tasks_count))) / @NumSeconds) * (CONVERT(DECIMAL(10,2),(SUM(w2.wait_time_ms) - SUM(w1.wait_time_ms))) / CONVERT(DECIMAL(10,2),(SUM(w2.waiting_tasks_count) - SUM(w1.waiting_tasks_count))))
    , @TotalWaitTimeInMSOverall = SUM(w2.wait_time_ms) - SUM(w1.wait_time_ms)
FROM (SELECT * FROM @ws ws1 WHERE ws1.RunNum = 1) w1
    INNER JOIN (SELECT * FROM @ws ws2 WHERE ws2.RunNum = 2) w2 ON w1.wait_type = w2.wait_type
WHERE (w2.waiting_tasks_count - w1.waiting_tasks_count) > 0;

SELECT @BatchReq2 = cntr_value
FROM sys.dm_os_performance_counters dopc
WHERE dopc.counter_name LIKE N'Batch Requests/sec%' COLLATE SQL_Latin1_General_CP1_CI_AS
    AND dopc.object_name = N'MSSQL$' + @InstanceName + N':SQL Statistics';

/*
    configured values for max server memory and min server memory, etc
*/
DECLARE @MaxServerMemory BIGINT;
DECLARE @MaxServerMemoryPages BIGINT;
DECLARE @MinServerMemory BIGINT;
DECLARE @MinPLE BIGINT;
DECLARE @RamMB BIGINT;
DECLARE @BufferPoolCommittedMB BIGINT;
DECLARE @BufferPoolCommitTargetMB BIGINT;
DECLARE @PercentOfDesiredSizeMB INT;
DECLARE @TargetPageLifeExpectancyPer4GB BIGINT;
SET @TargetPageLifeExpectancyPer4GB = 60 * 120; /* 120 minutes */
/*DECLARE @VMType VARCHAR(255);*/
DECLARE @PLESeconds BIGINT;

SELECT @MaxServerMemory = CONVERT(BIGINT,c.value)
FROM sys.configurations c
WHERE c.name = N'max server memory (mb)'

SET @MaxServerMemoryPages = @MaxServerMemory / 128; /* 8KB pages */

SELECT @MinServerMemory = CONVERT(BIGINT,c.value)
FROM sys.configurations c
WHERE c.name = N'min server memory (mb)'

SET @MinPLE = @MaxServerMemory / 4096E0 * @TargetPageLifeExpectancyPer4GB;

IF @VersionINT < 11
BEGIN
    SET @cmd = 'SELECT 
    @RamMB = dosi.physical_memory_in_bytes / 1048576
    , @BufferPoolCommittedMB = dosi.bpool_committed * 8192E0 / 1048576
    , @BufferPoolCommitTargetMB = dosi.bpool_commit_target * 8192E0 / 1048576
    , @PercentOfDesiredSizeMB = CONVERT(INT,(CONVERT(DECIMAL(18,2),dosi.bpool_committed) / dosi.bpool_commit_target) * 100)
FROM sys.dm_os_sys_info dosi;
';
END
ELSE 
BEGIN 
SET @cmd = 'SELECT 
    @RamMB = dosi.physical_memory_kb / 1024
    , @BufferPoolCommittedMB = dosi.committed_kb / 1024
    , @BufferPoolCommitTargetMB = dosi.committed_target_kb / 1024
    , @PercentOfDesiredSizeMB = CONVERT(INT,(CONVERT(DECIMAL(18,2),dosi.committed_kb) / dosi.committed_target_kb) * 100)
FROM sys.dm_os_sys_info dosi;';
END
EXEC sp_executesql @cmd
    , N'@RamMB BIGINT OUTPUT, @BufferPoolCommittedMB BIGINT OUTPUT, @BufferPoolCommitTargetMB BIGINT OUTPUT, @PercentOfDesiredSizeMB INT OUTPUT' 
    , @RamMB = @RamMB OUT
    , @BufferPoolCommittedMB = @BufferPoolCommittedMB OUT
    , @BufferPoolCommitTargetMB = @BufferPoolCommitTargetMB OUT
    , @PercentOfDesiredSizeMB = @PercentOfDesiredSizeMB OUT;

/*
    Page Life Expectancy for all memory nodes
*/
SELECT @PLESeconds = CONVERT(BIGINT, cntr_value) 
FROM sys.dm_os_performance_counters dopc
WHERE dopc.counter_name LIKE N'Page Life Expectancy%' COLLATE SQL_Latin1_General_CP1_CI_AS
    AND dopc.object_name = N'MSSQL$' + @InstanceName + N':Buffer Manager';

/*
    Total data in all user-databases.
*/
DECLARE @TotalDBSpaceUsed TABLE
(
    TotalSpaceUsedInMB BIGINT
);
DECLARE @SpaceUsedInMB BIGINT;
SET @cmd = '';
SELECT @cmd = @cmd + CASE WHEN @cmd = '' THEN '' ELSE '
UNION ALL
' END + 
'
SELECT DatabaseName = ''' + d.name + ''' 
    , AllocType = au.type_desc
    , TotalPagesInMB = SUM(au.total_pages) * 8192E0 / 1048576
FROM ' + QUOTENAME(d.name) + '.sys.allocation_units au
WHERE au.type > 0
GROUP BY au.type_desc
'
FROM master.sys.databases d
WHERE d.database_id > 4;
SET @cmd = 'SELECT SUM(TotalPagesInMB)
FROM (
' + @cmd + '
) t;'; 
INSERT INTO @TotalDBSpaceUsed (TotalSpaceUsedInMB)
EXEC sp_executesql @cmd;
SELECT @SpaceUsedInMB = TDSU.TotalSpaceUsedInMB
FROM @TotalDBSpaceUsed TDSU;

IF @Debug = 1
BEGIN
    SELECT ServerName = @@SERVERNAME
        , InstanceName = @InstanceName
        , DatabaseSpaceUsedMB = @SpaceUsedInMB
        , PLEinSeconds = @PLESeconds
        , MinAcceptablePLE = @MinPLE
        , MinServerMemoryMB = @MinServerMemory
        , MaxServerMemoryMB = @MaxServerMemory
        , TotalServerRAMinMB = @RamMB
        , BufferPoolCommittedMB = @BufferPoolCommittedMB
        , BufferPoolCommitTargetMB = @BufferPoolCommitTargetMB
        , PercentBufferPoolCommitted = @PercentOfDesiredSizeMB
        , BatchReqPerSecond = (@BatchReq2 - @BatchReq1) / @NumSeconds
        , LazyWritesPerSecond = (@LazyWrites2 - @LazyWrites1) / @NumSeconds
        , FreeListStallsPerSecond = (@FreeListStallsSec2 - @FreeListStallsSec2) / @NumSeconds
        /*, VMType = @VMType*/
        , IOTaskCount = @TaskCount 
        , TaskPerSecondAvg = @TasksPerSecondAvg 
        , AvgWaitTimeInMSPerTask = @AvgWaitTimeInMSPerTask 
        , AvgWaitTimeInMSPerSecond = @AvgWaitTimeInMSPerSecond 
        , TotalWaitTimeInMSOverall  = @TotalWaitTimeInMSOverall
        , SamplePeriodinSec = @NumSeconds;

    SELECT MaxServerMemorySuggested = 
            CASE WHEN @BufferPoolCommittedMB < @BufferPoolCommitTargetMB 
            THEN @BufferPoolCommittedMB 
            ELSE ((CONVERT(DECIMAL(18,4), @MinPLE) / @PLESeconds) * @MaxServerMemory) 
                    + (((@LazyWrites2 - @LazyWrites1) / @NumSeconds) * 64) 
                    + ((@FreeListStallsSec2 - @FreeListStallsSec2) / @NumSeconds) * 64 
            END
        , Reason = CASE WHEN @BufferPoolCommittedMB < @BufferPoolCommitTargetMB THEN N'Committed MB less than current Max Server Memory'
            ELSE N'Calculated based on PLE, Lazy Writes / second and List Stalls / second' END
        , LazyWritesX64 = (((@LazyWrites2 - @LazyWrites1) / @NumSeconds) * 64)
        , ListStallsX64 = ((@FreeListStallsSec2 - @FreeListStallsSec2) / @NumSeconds) * 64;
END

DECLARE @Out TABLE
(
    KeyID INT IDENTITY(1,1)
    , ItemDesc NVARCHAR(255)
    , ItemValue SQL_VARIANT
    , IsDebug BIT DEFAULT(0)
);

INSERT INTO @Out (ItemDesc, ItemValue, IsDebug)
VALUES (N'Server Name', CONVERT(NVARCHAR(255),@@SERVERNAME), 1);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Data Space Used (MB)', @SpaceUsedInMB);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Page Life Expectancy (sec)', @PLESeconds);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Minimum Acceptable Page Life Expectancy (sec)', @MinPLE);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Minimum Server Memory (MB)', @MinServerMemory);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Maximum Server Memory (MB)', @MaxServerMemory);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Total Server RAM in MB', @RamMB);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Buffer Pool Committed MB', @BufferPoolCommittedMB);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Buffer Pool Commit Target MB', @BufferPoolCommitTargetMB);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Percent of Buffer Pool Committed', @PercentOfDesiredSizeMB);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Batch Requests Per Second', (@BatchReq2 - @BatchReq1) / @NumSeconds);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Lazy Writes Per Second', (@LazyWrites2 - @LazyWrites1) / @NumSeconds);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Free List Stalls Per Second', (@FreeListStallsSec2 - @FreeListStallsSec2) / @NumSeconds);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'IO Task Count', @TaskCount);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Task Per Second Avg', @TasksPerSecondAvg);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Avg Wait Time In MS Per Task', @AvgWaitTimeInMSPerTask);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Avg Wait Time In MS Per Second', @AvgWaitTimeInMSPerSecond);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Total Wait Time In MS Overall', @TotalWaitTimeInMSOverall);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Sample Period in Seconds', @NumSeconds);

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Lazy Writes per Second', ((@LazyWrites2 - @LazyWrites1) / @NumSeconds));

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'List Stalls per Second', ((@FreeListStallsSec2 - @FreeListStallsSec2) / @NumSeconds));

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Recommended Max Memory (MB)', N'');

INSERT INTO @Out (ItemDesc, ItemValue)
VALUES (N'Recommended Max Memory Reason', N'');

INSERT INTO @Out (ItemDesc, ItemValue, IsDebug)
VALUES (N'Recommended Max Memory Signal', 0, 1);

/*
    Add memory if Lazy Writes occurred
    Add 64MB per Lazy Write (just for fun)
*/
DECLARE @LazyWritesMB INT;
SET @LazyWritesMB = (((@LazyWrites2 - @LazyWrites1) / @NumSeconds) * 64);

/*
    Add memory if Free List Stalls occurred
    Add 128MB per Free List Stall
*/
DECLARE @FreeListStallMB INT;
SET @FreeListStallMB = (((@FreeListStallsSec2 - @FreeListStallsSec2) / @NumSeconds) * 128);

/*
    Add the Additional memory requirements to the Recommended Max Memory row
*/
DECLARE @AdditionalMemory INT;
SET @AdditionalMemory = 
    @LazyWritesMB
    + @FreeListStallMB;

IF (@MaxServerMemory + @AdditionalMemory < 1024) AND (@PLESeconds >= @MinPLE)
BEGIN
    UPDATE @Out 
    SET ItemValue = @MaxServerMemory
    WHERE ItemDesc = N'Recommended Max Memory (MB)';

    UPDATE @Out 
    SET ItemValue = 'Max Server Memory is low, however PLE is acceptable'
    WHERE ItemDesc = N'Recommended Max Memory Reason';

    UPDATE @Out 
    SET ItemValue = 1
    WHERE ItemDesc = N'Recommended Max Memory Signal';
END

IF ((@BufferPoolCommittedMB + @AdditionalMemory) < @BufferPoolCommitTargetMB) AND (@PLESeconds >= @MinPLE)
BEGIN
    UPDATE @Out 
    SET ItemValue = @BufferPoolCommittedMB + @AdditionalMemory
    WHERE ItemDesc = N'Recommended Max Memory (MB)';

    UPDATE @Out 
    SET ItemValue = 'Buffer pool committed is less than Max Server Memory, and PLE is acceptable.'
    WHERE ItemDesc = N'Recommended Max Memory Reason';

    UPDATE @Out 
    SET ItemValue = 2
    WHERE ItemDesc = N'Recommended Max Memory Signal';
END

DECLARE @PLEMultiplier DECIMAL(10,2);
SET @PLEMultiplier = (CONVERT(DECIMAL(10,2),@MinPLE) / CONVERT(DECIMAL(10,2), @PLESeconds));
IF @PLEMultiplier < 0.90 SET @PLEMultiplier = 0.90;
IF @PLEMultiplier > 1.10 SET @PLEMultiplier = 1.10;

INSERT INTO @Out (ItemDesc, ItemValue, IsDebug)
VALUES (N'PLE Multiplier', @PLEMultiplier, 1);

IF /*(@MaxServerMemory + @AdditionalMemory >= 1024) AND*/ (@PLESeconds <= @MinPLE)
BEGIN
    UPDATE @Out 
    SET ItemValue = 
        (SELECT TOP(1) Inc
        FROM (
            SELECT Inc = t.RowNum * 256
            FROM (
                SELECT RowNum = CONVERT(BIGINT,ROW_NUMBER() OVER (ORDER BY o.object_id))
                FROM sys.objects o, sys.objects o1
                ) t
            WHERE (t.RowNum * 256) <  CONVERT(BIGINT,POWER(2,30))
            ) t1
        WHERE t1.Inc > CONVERT(INT, (@MaxServerMemory * @PLEMultiplier))
        ORDER BY t1.Inc)
    WHERE ItemDesc = N'Recommended Max Memory (MB)';

    UPDATE @Out 
    SET ItemValue = 'Low PLE indicates Max Server Memory should be adjusted upwards.'
    WHERE ItemDesc = N'Recommended Max Memory Reason';

    UPDATE @Out 
    SET ItemValue = 3
    WHERE ItemDesc = N'Recommended Max Memory Signal';
END

IF (@MaxServerMemory + @AdditionalMemory >= 1024) AND (@PLESeconds > @MinPLE)
BEGIN
    UPDATE @Out 
    SET ItemValue = 
        (SELECT TOP(1) Inc
        FROM (
            SELECT Inc = t.RowNum * 256
            FROM (
                SELECT RowNum = CONVERT(BIGINT,ROW_NUMBER() OVER (ORDER BY o.object_id))
                FROM sys.objects o, sys.objects o1
                ) t
            WHERE (t.RowNum * 256) <  CONVERT(BIGINT,POWER(2,30))
            ) t1
        WHERE t1.Inc <= CONVERT(INT, (@MaxServerMemory * @PLEMultiplier))
        ORDER BY t1.Inc DESC)
    WHERE ItemDesc = N'Recommended Max Memory (MB)';

    UPDATE @Out 
    SET ItemValue = 'High PLE indicates Max Server Memory could be adjusted downwards.'
    WHERE ItemDesc = N'Recommended Max Memory Reason';

    UPDATE @Out 
    SET ItemValue = 4
    WHERE ItemDesc = N'Recommended Max Memory Signal';
END

DECLARE @RecommendedMaxServerMemory INT;
SELECT  @RecommendedMaxServerMemory = CONVERT(INT,ItemValue)
FROM @Out o 
WHERE o.ItemDesc = N'Recommended Max Memory (MB)';

IF @RecommendedMaxServerMemory > (@MaxServerMemory * 0.96)
    AND @RecommendedMaxServerMemory < (@MaxServerMemory * 1.04)
BEGIN
    UPDATE @Out
    SET ItemValue = @MaxServerMemory
    WHERE ItemDesc = N'Recommended Max Memory (MB)';
    UPDATE @Out
    SET ItemValue = 'No changed recommended'
    WHERE ItemDesc = N'Recommended Max Memory Reason';
    UPDATE @Out 
    SET ItemValue = 0
    WHERE ItemDesc = N'Recommended Max Memory Signal';
END 

IF (@HTMLOutput = 1)
BEGIN
    SELECT ItemValue
        , HTMLOutput = '<table>' + 
            (
                SELECT 'td' = ItemDesc
                    , ''
                    , 'td' = ItemValue
                    , ''
                FROM @Out o
                WHERE CASE WHEN @Debug = 0 THEN o.IsDebug ELSE 0 END = 0
                ORDER BY o.KeyID
                FOR XML PATH('tr')
            ) +
            '</table>'
    FROM @Out o
    WHERE o.ItemDesc = N'Recommended Max Memory Signal';
END
ELSE
BEGIN
    SELECT *
    FROM @Out o
    WHERE CASE WHEN @Debug = 0 THEN o.IsDebug ELSE 0 END = 0
    ORDER BY o.KeyID;
END

This code compares the PLE against a minimum "acceptable" PLE for the amount of max server memory the system has configured. If the PLE is appreciably higher than the acceptable number, it suggests a maximum of 10% lower max server memory. If the PLE is lower than the acceptable PLE, it suggests a maximum of 10% more max server memory.

If the actual amount of committed buffer pool is less than the target buffer pool size it suggests lowering max server memory to that amount, plus some extra memory for threads, lazy writes, etc.

The code also looks at various performance counters for things like Lazy Writes / second, Free List Stalls, and Batch Requests.

The code is not perfect, I'm sharing it here to get input, and for the benefit of future SO users.

  • 1
    Max Mind you starting from SQL Server 2012 Buffer Pool target and committed has no sense and these counters are deprecated. Instead you must use Memory Manager Target Committed (KB) and current committed. If you want to read more why it gives wrong value social.technet.microsoft.com/wiki/contents/articles/… – Shanky Dec 13 '14 at 23:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.