Balancing CPU and I/O throughput is essential to achieve good overall performance and to maximize hardware utilization.

SQL Server includes two asynchronous I/O mechanisms - sequential read ahead and random prefetching - that are designed to address this challenge.

To understand why asynchronous I/O is so important, consider the CPU to I/O performance gap. The memory subsystem on a modern CPU can deliver data sequentially at roughly 5 Gbytes per second per socket (or for non-NUMA machines for all sockets sharing the same bus) and (depending on how you measure it) can fetch random memory locations at roughly 10 to 50 million accesses per second. By comparison, a high end 15K SAS hard drive can read only 125 Mbytes per second sequentially and can perform only 200 random I/Os per second (IOPS). Solid State Disks (SSDS) can reduce the gap between sequential and random I/O performance by eliminating the moving parts from the equation, but a performance gap remains. In an effort to close this performance gap, it is not uncommon for servers to have a ratio of 10 or more drives for every CPU. (It is also important to consider and balance the entire I/O subsystem including the number and type of disk controllers not just the drives themselves but that is not the focus of this post.)

Unfortunately, a single CPU issuing only synchronous I/Os can keep only one spindle active at a time. For a single CPU to exploit the available bandwidth and IOPs of multiple spindles effectively the server must issue multiple I/Os asynchronously. Thus, SQL Server includes the aforementioned read ahead and prefetching mechanisms. In this post, I‘ll take a look at sequential read ahead.

When SQL Server performs a sequential scan of a large table, the storage engine initiates the read ahead mechanism to ensure that pages are in memory and ready to scan before they are needed by the query processor. The read ahead mechanism tries to stay 500 pages ahead of the scan. SQL Server tries to combine up to 64 contiguous pages (512 Kbytes) into a single  scatter (asynchronous) I/O.  So, in a best case scenario, it can read ahead 500 pages in just 8 I/Os.  However, if the pages in the table are not contiguous (e.g., due to fragmentation), SQL Server cannot combine the I/Os and must issue one I/O per page (8 Kbytes).

We can see the read ahead mechanism in action by checking the output of SET STATISTICS IO ON. For example, I ran the following query on a 1GB scale factor TPC-H database. The LINEITEM table has roughly 6 million rows.

SET STATISTICS IO ON

SELECT COUNT(*) FROM LINEITEM

Table ‘LINEITEM‘. Scan count 3, logical reads 22328, physical reads 3, read-ahead reads 20331, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Repeating the query a second time shows that the table is now cached in the buffer pool:

SELECT COUNT(*) FROM LINEITEM

Table ‘LINEITEM‘. Scan count 3, logical reads 22328, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

For sequential I/O performance, it is important to distinguish between allocation ordered and index ordered scans.

An allocation ordered scan tries to read pages in the order in which they are physically stored on disk while an index ordered scan reads pages according to the order in which the data on those index pages is sorted. (Note that in many cases there are multiple levels of indirection such as RAID devices or SANS between the logical volumes that SQL Server sees and the physical disks. Thus, even an allocation ordered scan may in fact not be truly optimally ordered.)

Although SQL Server tries to sort and read pages in allocation order even for an index ordered scan, an allocation ordered scan is generally going to be faster since pages are read in the order that they are written on disk with the minimal number of seeks. Heaps have no inherent order and, thus, are always scanned in allocation order. Indexes are scanned in allocation order only if the isolation level is read uncommitted (or the NOLOCK hint is used) and only if the query process does not request an ordered scan.Defragmenting indexes can help to ensure that index ordered scans perform on par with allocation ordered scans.

Correction:

First - "The memory subsystem on a modern CPU can deliver data sequentially at roughly 5 Gbytes per second per core"

There are NUMA and non-NUMA hardware systems. Non-NUMA share FSB for exclusive access to RAM, so it becomes 5 Gbytes perl ALL CPUs on board. For NUMA systesms - all cores on the same NUMA node share that 5Gbytes. Thus its not per core.

Second is "Solid State Disks (SSDS) can reduce the gap between sequential and random I/O performance by eliminating the moving parts from the equation, but a performance gap remains." Do you mean performance = throughput here?

Thank you, Serge

The non-leaf nodes of the B-tree (specifically those nodes one level above the leaf nodes) have pointers to and can be used to prefetch multiple pages at a time.  Of course, this optimization does work for heaps which do not have non-leaf nodes.