Advances in Memory Management for Windows
Download 92.35 Kb.
|
- Navigate this page:
- NUMA-Aware System Functions for Applications
- NUMA-Aware System Functions for Drivers
- Efficiency and Parallelism
- Page-Frame Number and PFN Database
- Cache-Aligned Pool Allocation
- Additional Optimizations
- Diagnosis of Hardware Errors
- Code Integrity and Driver Signing
- Data Preservation during Bug Checks
- For Hardware Manufacturers
- For Application Developers
- For System Administrators
- Windows Hardware and Driver Central
- Microsoft TechNet: Inside the Windows Vista Kernel: Part 3 (April 2007)
Interrupt AffinityDrivers for PCI devices that support message-signaled interrupts (MSI or MSI-X) can specify an interrupt affinity—that is, the set of processors on which the device’s interrupt service routine (ISR) runs—for each MSI message that the device generates. This feature can significantly improve performance, especially on network interface cards (NICs) that support receive-side scaling (RSS). A driver can specify an affinity for a particular MSI message when it connects the interrupt. A driver can also set default affinity and affinity policy by setting values for the Interrupt Management\Affinity Policy registry key in the DDInstall.HW section of its INF. An administrator can also set these values in the registry. For more information, see the WinHEC presentation "NUMA I/O Optimizations," the white paper "Interrupt Architecture Enhancements in Windows," and "Interrupt Affinity and Priority" in the WDK. For NIC-specific information, see "NDIS MSI-X" in the WDK.
PagingWindows Server 2008 incorporates further NUMA enhancements for paging. Server 2008 prefetches pages to the application’s ideal node and migrates pages to the ideal node when a soft page fault occurs. A soft page fault occurs when the system can find the requested page elsewhere in memory, whereas a hard page fault requires reading the page in from disk. ScalabilityAs Windows runs on larger and more powerful machines, Microsoft continues to enhance the system’s ability to scale up to service more and faster processors and RAM. Efficiency and ParallelismAs a result of numerous internal improvements to the memory manager, memory allocation now requires fewer I/O operations and fewer locks for optimal throughput. Internally, the memory manager now uses a bitmap instead of a linked list to track the free pages in the nonpaged pool. Unlike a linked list, a bitmap can be searched without a lock, thus reducing contention for the associated lock by more than 50 percent. Furthermore, bitmaps provide automatic coalescing of contiguous free pages. Windows Server 2008 also uses bitmaps to describe system PTEs. Large shared sets are now directly mapped instead of hashed. When the number of entries in a hash table is very large, frequent collisions typically occur unless the hash table can be dynamically resized. However, resizing is expensive to perform on large sets. Direct mapping is therefore more efficient than hashing in this situation. In Windows Server 2008, the allocation of physically contiguous memory is greatly enhanced. Requests to allocate contiguous memory are much more likely to succeed because the memory manager now dynamically replaces pages, typically without trimming the working set or performing I/O operations. In addition, many more types of pages—such as kernel stacks and file system metadata pages, among others—are now candidates for replacement. Consequently, more contiguous memory is generally available at any given time. In addition, the cost to obtain such allocations is greatly reduced. Page-Frame Number and PFN DatabaseThe page-frame number (PFN) database contains information about all of the physical memory in the machine. In 64-bit editions of Windows Vista SP1 and Windows Server 2008, page-frame numbers are 64 bits long to support large amounts of memory and NUMA architectures, on which the physical address space is sometimes sparsely populated with memory. In earlier Windows releases, whenever a new page was needed, the memory manager acquired the PFN spinlock and removed a new page from the appropriate list chained through the PFN database. Instead, Windows Vista maintains short lists of immediately available zero pages and free pages for each NUMA node and page color. (Page coloring is a technique that the memory manager uses to reduce the possibility of cache-line collisions among pages.) In many cases—particularly demand-zero faults and copy-on-write faults—the system can now get a single page without acquiring the PFN lock. Reducing the number of spin lock acquisitions eliminates potential spins on other processors and thus improves parallelism. Large PagesWindows Server 2003 introduced large pages for user-mode applications. Windows Vista and Windows Server 2008 use large pages more extensively internally and provide enhanced support for them. Windows Vista and Windows Server 2008 use large pages for the following: Initial nonpaged pool PFN database User application and driver images Page file-backed shared memory User-mode VirtualAlloc allocations Driver I/O space mappings
Table 1. Large Page Sizes
An application can call GetLargePageMinimum to determine the current large page size. The Windows Vista memory manager allocates ranges of large pages more quickly than earlier Windows releases did. The entire range is no longer required to be contiguous, so that attempts to allocate large pages are more likely to succeed and less likely to cause page thrashing. For example, if an application requests 10 MB of large pages, Windows Vista and later Windows releases can allocate five large pages of 2 MB each (if large pages are 2 MB on the individual hardware platform) instead of trying to find 10 MB of physically contiguous memory. The Windows Vista memory manager also keeps track of which NUMA nodes the allocated memory belongs to and can zero large pages in parallel by dispatching threads to the appropriate nodes to zero them locally. Cache-Aligned Pool AllocationWindows Vista implements support for cache-aligned pool allocation. Drivers can specify the following flags in the ExAllocatePoolXxx functions to request cache-aligned memory: NonPagedPoolCacheAligned PagedPoolCacheAligned These flags were defined in earlier Windows releases, but they were ignored. Virtual MachinesEfficiency and scalability are not only important for good server performance, they are also critical for Windows to run as a guest operating system in a virtualized system. Windows Vista incorporates several changes that improve its performance in virtual machine scenarios. The translation look-aside buffer (TLB) caches the translation from VAs to physical addresses so that the processor can quickly access this information. If an address is not available in the TLB, the processor typically must make several memory references, which are quite time consuming. Consequently, overall system performance decreases as the TLB hit rate drops. One way to increase the likelihood that an address will be in the TLB is to flush the TLB less often. Each time a page has been made invalid, its entry must be flushed from the buffer. A page becomes invalid when it is unmapped, freed, trimmed from the working set, or modified by a copy-on-write operation, among others. The entry must also be flushed if changes are made to the protection or cache attributes of the page. Flushing the entire translation buffer across all processors is a relatively expensive operation that requires significant operating system overhead. Furthermore, after the buffers are flushed, they must be repopulated. Windows Vista rarely flushes the entire buffer. As a result, virtual machines can operate much more efficiently. If a virtualized system hosts several guest operating systems, the size of the guests can constrain hypervisor performance and limit scalability. To use a smaller memory footprint and thus be a better guest system, Windows Server 2008 frees unneeded memory that it has speculatively allocated. In particular, the system reclaims memory from the initial nonpaged pool if it is not being used.
Load BalancingWindows Vista exports the following new events to help in load balancing: LowCommitConditionNotification HighCommitConditionNotification MaximumCommitConditionNotification The LowCommitConditionNotification event is set when the operating system's commit charge is low, relative to the current commit limit. In other words, memory usage is low and a lot of space is available for allocations. The HighCommitConditionNotification event is set when the operating system's commit charge is high, relative to the current commit limit. In other words, memory usage is high and very little space is available. If adequate disk space is available, the system obtains more memory by automatically increasing the page file size up to the limit imposed by the administrator. A short-term option is to reduce the current system load. Long-term solutions are to increase the minimum page file size or add RAM. The MaximumCommitConditionNotification event is set when the operating system's commit charge is near the maximum commit limit. In other words, memory usage is very high, very little space is available, and the system cannot increase the size of its paging files because of the current limits imposed by the administrator. A system administrator can always increase the size or number of paging files, without restarting the computer, if adequate disk space is available. Other alternatives are to increase the minimum or maximum page file sizes, add RAM, or reduce the load. These events supplement the pool notification events that were added in Windows Server 2003. Drivers and other kernel-mode components can register for these events. For more information on memory-related notification events, see "Standard Event Objects" in the WDK. Additional OptimizationsAdditional memory manager optimizations involve the following areas: VirtualAlloc and address windowing extensions (AWE) allocations. VirtualProtect function. Windows on Windows (WOW) on 64-bit systems. Windows acquires a per-process address space lock to synchronize changes to the user address space. In Windows Vista, this lock supports both shared and exclusive access, whereas in earlier Windows versions, the lock supported exclusive access only. Consequently, many operations such as VirtualAlloc and VirtualQuery can now run in parallel. Overall, changes within VirtualAlloc reduce the time required for AWE allocations by over 2500 percent in some scenarios. The VirtualProtect function changes the access protection on a region of pages in virtual memory. When a page’s access protection attribute changes, processors must flush the corresponding TLB entry. Windows Server 2008 issues a single flush request to all processors whose TLB might contain the entry instead of multiple single requests to each individual processor. As a result, VirtualProtect can change access protection on large regions 60 times faster than in earlier Windows versions. On 64-bit architectures, Windows Vista uses demand-zeroed memory instead of pool memory to allocate the page-table bitmaps for 32-bit binary emulation. This change enables 32-bit binaries to run much more efficiently because they require a smaller system memory footprint and perform fewer I/O operations.
System IntegrityThrough online crash analysis (OCA), users can upload data about system crashes to Microsoft. This data has provided useful information about the causes of common system crashes and has led to several system enhancements to detect and handle potential system corruption. Windows Vista and Windows Server 2008 incorporate advances to improve system integrity in the following areas: Diagnosis of hardware errors Code integrity and driver signing Data preservation during bug checks Diagnosis of Hardware ErrorsAs mentioned in "Page-File Writes" earlier in this paper, Windows maintains a list of zero pages. Hardware errors such as DMA transfer errors and single bit errors can corrupt memory after the pages have initially been zeroed, so Windows Vista checks the list to ensure that these pages actually are zero. If the system finds an error, it records the physical address at which the error occurred and the nature of the error in the event log. This information helps to pinpoint single-bit errors that are caused by hardware failures. Machines that frequently encounter such errors are often prone to application hangs and crashes that are extremely difficult to track down. OCA data indicates that such failures are much more common than previously suspected. Independent hardware vendors (IHVs) can help diagnose and prevent such errors by using error-correcting code (ECC) memory. Code Integrity and Driver SigningThe memory manager implements a simple, high-speed technique to validate images for code integrity. This feature enforces the mandatory code signing for kernel-mode drivers on x64-based systems. For more information on code signing, see "Digital Signatures for Kernel Modules on x64-based Systems Running Windows Vista," "Kernel-Mode Code Signing Walkthrough," and "Summary of Windows Driver Signing Requirements." Windows Vista supports hot-patching for both system-wide and session drivers, so that patches can be installed without rebooting the user’s system. Thus, users can take advantage of security patches as soon as they become available without waiting for reboot.
To further protect system data, Windows Vista supports the ability to mark views of the system cache as read only. The registry uses this feature to protect its views from inadvertent driver corruption. Thus, registry data is read only except when it is actively being modified. Driver writers can use the new .pagein debugger command to view the contents of kernel-mode memory addresses that have been paged out to disk. For more information about this command, see Debugging Tools for Windows.
What You Should DoMost of the memory management enhancements that are described in this paper are internal and are transparent to administrators, software developers, and hardware manufacturers. However, a few of the changes require awareness or action to gain maximum benefit and to contribute to an improved user experience. Here are the most important effects for hardware manufacturers, driver developers, application developers, and system administrators. For Hardware ManufacturersUse ECC. For Driver DevelopersNever attempt to access memory beyond what the driver has allocated. Use Driver Verifier to catch this error. Handle dummy pages correctly in drivers that directly access the contents of MDLs. Use KeExpandKernelStackAndCallout as necessary to gain additional kernel stack space. Use the new NUMA-aware system functions in drivers that are sensitive to NUMA architectures and specify interrupt affinity if this is important for your device. Use new events for notification about system load if your driver allocates memory that it can release during operation. Be aware of driver signing requirements for Windows Vista, particularly for x64 architectures. Use the .pagein debugger command to inspect kernel-mode data that was paged out.
Be aware that the default NUMA node for memory allocation is now the ideal node instead of the current node. Use the new NUMA-aware system functions to control memory allocation and query page locations on NUMA architectures.
Use a debugger with the !vm 21 command to inspect details of kernel VA space use on 32-bit systems. Check the system event log for zero-page corruption errors. Upload crash data to OCA whenever possible.
ResourcesMSDN:Windows Vista ISV Security http://msdn2.microsoft.com/en-us/library/bb430720.aspx Memory Management Registry Keys http://msdn2.microsoft.com/en-us/library/bb870880.aspx Windows Driver Kit: http://msdn2.microsoft.com/en-us/library/aa972908.aspx Kernel-Mode Driver Architecture Design Guide Memory Management Interrupt Affinity and Priority Standard Event Objects Kernel-Mode Driver Architecture Reference Standard Driver Routines Driver Support Routines Driver Development Tools Boot Options for Driver Testing and Debugging Network Design Guide NDIS MSI-X Windows Hardware and Driver Central:Driver Signing Requirements for Windows [home page] http://www.microsoft.com/whdc/winlogo/drvsign/drvsign.mspx Digital Signatures for Kernel Modules on Systems Running Windows Vista Summary of Windows Kernel-Mode Driver Signing Requirements Windows PC Accelerators: Performance Technology for Windows Vista http://www.microsoft.com/whdc/system/sysperf/accelerator.mspx What Is Really in That MDL? http://www.microsoft.com/whdc/driver/tips/mdl.mspx Interrupt Architecture Enhancements in Windows http://www.microsoft.com/whdc/system/bus/PCI/MSI.mspx NUMA I/O Optimizations http://download.microsoft.com/download/a/f/d/afdfd50d-6eb9-425e-84e1-b4085a80e34e/SVR-T332_WH07.pptx Microsoft TechNet:Inside the Windows Vista Kernel: Part 3 (April 2007) http://www.microsoft.com/technet/technetmag/issues/2007/04/VistaKernel/ Book:Windows Internals, Fourth Edition, Russinovich, Mark, and David A. Solomon. Redmond, WA: Microsoft Press, 2005 http://www.microsoft.com/MSPress/books/6710.aspx Directory: download download -> Performance Tuning Guidelines for Windows Server 2008 May 20, 2009 Abstract download -> Northern England’s set-jetting locations download -> Bbc archives: Beyond the programme download download -> Occupational health and safety download -> Dynatest 3031 lwd light Weight Deflectometer download -> Forth coming competitive exams download -> English Olympiad Tasks: 10 th form Аудіювання, 10 клас. Текст download -> Brush High School Morning Announcements Friday, September 05, 2014 all school download -> Year 9 The Arts — Music download -> Years 7 and 8 standard elaborations — Australian Curriculum: Music draft Download 92.35 Kb. Share with your friends:
|