Core Application Infrastructure Services

Performance Optimizations

• 
  New Winsock driver (AFD) that provides the capability to complete large TransmitFile operations in an APC of the transmitting thread rather than posting it to a delayed system worker thread.
  
• 
  Enhanced memory allocation – the per-processor look-aside lists reduce shared memory access of global look-aside list headers leading to a speedup of five percent for disk I/O.
  
• 
  Additional non-paged and paged pool lists reduce pool fragmentation.
  
• 
  Reduced hold time for the dispatch lock on workloads such as the TPC-C (such as Microsoft SQL Server 7.0 with fibers) reduces contention on key system resources by up to 30 percent.
  
• 
  Fibers in Microsoft SQL Server 7.0 reduce context switches and improve throughput by up to 18 percent compared with threads.
  
• 
  Increased maximum working set for the file system cache from 512 MB in Windows NT Server 4.0 to 960 MB in Windows 2000 Server. This reduces contention on system resources, thereby reducing context switching. On a SpecWeb ’96 workload, this improves throughput by up to 5 percent.
  
• 
  Per-processor completion ports reduces CPU migration of threads, which provides a 5 percent to 7 percent increase in TPC-C throughput with Microsoft SQL Server 6.5.
  
• 
  NTFS improvements reduce the number of operations posted to system worker threads, reducing context switching and thread CPU migrations by 46 percent and improving dual processor throughput on NetBench 5.0 by up to 3 percent.
  
• 
  Increased use of shared locks for NTFS TransactionTable executive resource reduces contention on this resource by up to 14 percent for a multi-processor file-server workload.
  
• 
  Reduced SCSI miniport controller contention (by a magnitude of seven) on Windows 2000 Server compared to Windows NT Server 4.0 on a TPC-C workload with Microsoft SQL Server 7.0.
  
• 
  Interrupt affinity provides an improvement of up to 7 percent on a SpecWeb ’96 workload on a four-processor system using four NICs.
  
• 
  Reduction of TCP/IP contention is expected to improve four-processor SpecWeb ’96 scaling by up to 20 percent.
  
• 
  Support for I2O hardware has been added. With I2O, several significant performance enhancing benefits can be achieved. The most significant is the offloading of certain I/O operations to intelligent storage adapters, resulting in more available CPU cycles to process complex calculations and lower overall CPU usage.
  

Other Core Enhancements

• 
  Job Objects are new kernel objects that can be named and secured. They are used to collect a group of related processes, enabling management and tracking of the process group. The Job Object enforces job quotas and security context. This enables the monitoring and control of per-process CPU time, per-job CPU time, minimum and maximum working set (memory usage), active process count, CPU affinity (which CPUs in a multi-processor system can run the processes) and priority class.
  
• 
  Scatter/Gather I/O support enables higher I/O throughput when application data is located in non-contiguous memory locations (which is typical) and data needs to be written to a contiguous file location. It is VLM-enabled on the Alpha platform in Windows 2000 Server. The WriteFileGather API takes pointers to one or more pages in memory, gathers them together, and writes them out to the file as one chunk. ReadFileScatter reads in one or more pages from the file system and scatters them to pre-established buffers. The advantage of this technique is that the program need not work with intermediate buffers that contain the data as a single logical chunk.
  
• 
  Spin Count is a new feature introduced to deal with the bottlenecks created by a specific memory block or resource that is constantly being acquired or released by a particular process. Usually a critical section guards these resources. However, if a thread blocks on a critical section, it is calling WaitForSingleObject(), which is relatively expensive. If the critical section for a resource is usually acquired and released in a fairly short time period, the critical section can be optimized so that threads will not spend as much time in an expensive WaitForSingleObject() call. The dwReserved field of a critical section is now used for a spin count. If the spin count is set, a thread that would normally block while waiting for a critical section will instead enter a loop, where it continually checks to see if the critical section can be acquired. If the loop executes “spin count” a number of times, the thread gives up and reverts back to the old behavior by calling WaitForSingleObject(). The goal is that the blocking processor should acquire the critical section faster by this method than by using WaitForSingleObject(). It should be noted that spin counts do not result in any performance gain on single processor systems, but that the APIs can be called with no ill effects.
  
• 
  High Performance Sorting support has been introduced in Windows 2000 Advanced Server. Specifically, commercial sorting performance of large data sets have been optimized, improving the performance of common tasks such as preparing to load data in batch for data warehouse/data mart operations and to prepare large sort-sensitive print and batch operations.