Document revision date: 19 July 1999

The following publications provide more information on local area networks.

• IEEE Standards 802.1 (A, B, C, and D), 802.2, and 802.3.
• The Ethernet--Data Link Layer and Physical Layer Specification
• ANSI X3t9.5 and X3.139
• Digital FDDI Network Architecture

This chapter includes updated information for OpenVMS Version 7.2.

As of Version 7.0, OpenVMS Alpha includes two features that provide dramatically improved I/O performance: Fast I/O and Fast Path. These features are designed to promote OpenVMS as a leading platform for database systems. Performance improvement results from reducing the CPU cost per I/O request and improving symmetric multiprocessing (SMP) scaling of I/O operations. The CPU cost per I/O is reduced by optimizing code for high-volume I/O and by using better SMP CPU memory cache. SMP scaling of I/O is increased by reducing the number of spinlocks taken per I/O and by substituting finer-granularity spinlocks for global spinlocks.

The improvements follow a natural division that already exists between the device-independent and device-dependent layers in the OpenVMS I/O subsystem. The device-independent overhead is addressed by Fast I/O, which is a set of lean system services that can substitute for certain $QIO operations. Using these services requires some coding changes in existing applications, but the changes are usually modest and well contained. The device-dependent overhead is addressed by Fast Path, which is an optional performance feature that creates a "fast path" to the device. It requires no application changes.

Fast I/O and Fast Path can be used independently. However, together they can provide a 45% reduction in CPU cost per I/O on uniprocessor systems and a 52% reduction on multiprocessor systems.

10.1 Fast I/O

Fast I/O is a set of three system services that were developed as a $QIO alternative built for speed. These services are not a $QIO replacement; $QIO is unchanged, and $QIO interoperation with these services is fully supported. Rather, the services substitute for a subset of $QIO operations, namely, only the high-volume read/write I/O requests.

The Fast I/O services support 64-bit addresses for data transfers to and from disk and tape devices.

While Fast I/O services are available on OpenVMS VAX, the performance advantage applies only to OpenVMS Alpha. OpenVMS VAX has a run-time library (RTL) compatibility package that translates the Fast I/O service requests to $QIO system service requests, so one set of source code can be used on both VAX and Alpha systems.

10.1.1 Fast I/O Benefits

The performance benefits of Fast I/O result from streamlining high-volume I/O requests. The Fast I/O system service interfaces are optimized to avoid the overhead of general-purpose services. For example, I/O request packets (IRPs) are now permanently allocated and used repeatedly for I/O rather than allocated and deallocated anew for each I/O.

The greatest benefits stem from having user data buffers and user I/O status structures permanently locked down and mapped using system space. This allows Fast I/O to do the following:

• For direct I/O, avoid per-I/O buffer lockdown or unlocking.
• For buffered I/O, avoid allocation and deallocation of a separate system buffer, since the user buffer is always addressable.
• Complete Fast I/O operations at IPL 8, thereby avoiding the interrupt chaining usually required by the more general-purpose $QIO system service. For each I/O, this eliminates the IPL 4 IOPOST interrupt and a kernel AST.

In total, Fast I/O services eliminate four spinlock acquisitions per I/O (two for the MMG spinlock and two for the SCHED spinlock). The reduction in CPU cost per I/O is 20% for uniprocessor systems and 10% for multiprocessor systems.

10.1.2 Using Buffer Objects

The lockdown of user-process data structures is accomplished by buffer objects. A "buffer object" is process memory whose physical pages have been locked in memory and double-mapped into system space. After creating a buffer object, the process remains fully pageable and swappable and the process retains normal virtual memory access to its pages in the buffer object.

If the buffer object contains process data structures to be passed to an OpenVMS system service, the OpenVMS system can use the buffer object to avoid any probing, lockdown, and unlocking overhead associated with these process data structures. Additionally, double-mapping into system space allows the OpenVMS system direct access to the process memory from system context.

To date, only the $QIO system service and the Fast I/O services have been changed to accept buffer objects. For example, a buffer object allows a programmer to eliminate I/O memory management overhead. On each I/O, each page of a user data buffer is probed and then locked down on I/O initiation and unlocked on I/O completion. Instead of incurring this overhead for each I/O, it can be done once at buffer object creation time. Subsequent I/O operations involving the buffer object can completely avoid this memory management overhead.

Two system services can be used to create and delete buffer objects, respectively, and can be called from any access mode. To create a buffer object, the $CREATE_BUFOBJ system service is called. This service expects as inputs an existing process memory range and returns a buffer handle for the buffer object. The buffer handle is an opaque identifier used to identify the buffer object on future I/O requests. The $DELETE_BUFOBJ system service is used to delete the buffer object and accepts as input the buffer handle. Although image rundown deletes all existing buffer objects, it is good form for the application to clean up properly.

A 64-bit equivalent version of the $CREATE_BUFOBJ system service ($CREATE_BUFOBJ_64) can be used to create buffer objects from the new 64-bit P2 or S2 regions. The $DELETE_BUFOBJ system service can be used to delete 32-bit or 64-bit buffer objects.

Buffer objects require system management. Because buffer objects tie up physical memory, extensive use of buffer objects require system management planning. All the bytes of memory in the buffer object are deducted from a systemwide SYSGEN parameter called MAXBOBMEM (maximum buffer object memory). System managers must set this parameter correctly for the application loads that run on their systems.

The MAXBOBMEM parameter defaults to 100 Alpha pages, but for applications with large buffer pools it will likely be set much larger. To prevent user-mode code from tying up excessive physical memory, user-mode callers of $CREATE_BUFOBJ must have a new system identifier, VMS$BUFFER_OBJECT_USER, assigned. This new identifier is automatically created in an OpenVMS Version 7.0 upgrade if the file SYS$SYSTEM:RIGHTSLIST.DAT is present. The system manager can assign this identifier with the DCL command SET ACL command to a protected subsystem or application that creates buffer objects from user mode. It may also be appropriate to grant the identifier to a particular user with the Authorize utility command GRANT/IDENTIFIER (for example, to a programmer who is working on a development system).

There is currently a restriction on the type of process memory that can be used for buffer objects. Global section memory cannot be made into a buffer object.

10.1.3 Differences Between Fast I/O Services and $QIO

The precise definition of high-volume I/O operations optimized by Fast I/O services is important. I/O that does not comply with this definition either is not possible with the Fast I/O services or is not optimized. The characteristics of the high-volume I/O optimized by Fast I/O services can be seen by contrasting the operation of Fast I/O system services to the $QIO system service as follows:

• The $QIO system service I/O status block (IOSB) is replaced by an I/O status area (IOSA) that is larger and quadword aligned. The transfer byte count returned in IOSA is 64 bits, and the field is aligned on a quadword boundary. Unlike the IOSB, which is optional, the IOSA is required.
• User data buffers must be aligned to a 512-byte boundary.
• All user process structures passed to the Fast I/O system services must reside in buffer objects. This includes the user data buffer and the IOSA.
• Only transfers that are multiples of 512 bytes are supported.
• Only the following function codes are supported: IO$_READVBLK, IO$_READLBLK, IO$_WRITEVBLK, and IO$_WRITELBLK.
• Only I/O to disk and tape devices is optimized for performance.
• No event flags are used with Fast I/O services. If application code must use an event flag in relation to a specific I/O, then the Event No Flag EFN (EFN$C_ENF) can be used. This event flag is a no-overhead EFN that can be used in situations when an EFN is required by a system service interface but has no meaning to an application.
  For example, Fast I/O services do not use EFNs, so the application cannot specify a valid EFN associated with the I/O to the $SYNCH system service with which to synchronize I/O completion. To resolve this issue, the application can call the $SYNCH system service passing as arguments: EFN$C_ENF and the address of the appropriate IOSA. Specifying EFN$C_ENF signifies to $SYNCH that no EFN is involved in the synchronization of the I/O. Once the IOSA has been written with a status and byte count, return from the $SYNCH call occurs. The IOSA is now the central point of synchronization for a given Fast I/O (and is the only way to determine whether the asynchronous I/O is complete).
• To minimize argument passing overhead to these services, the $QIO parameters P3 through P6 are replaced by a single argument that is passed directly by the Fast I/O system services to device drivers. For disk-like devices, this argument is the media address (VBN or LBN) of the transfer. For drivers with complex parameters, this argument is the address of a descriptor or of a buffer specific to the device and function.
• Segmented transfers are supported by Fast I/O but are not fully optimized. There are two major causes of segmented transfers. The first is disk fragmenting. While this can be an issue, it is assumed that sites seeking maximum performance have eliminated the overhead of segmenting I/O due to fragmentation.
  A second cause of segmenting is issuing an I/O that exceeds the port's maximum limit for a single transfer. Transfers beyond the port maximum limit are segmented into several smaller transfers. Some ports limit transfers to 64K bytes. If the application limits its transfers to less than 64K bytes, this type of segmentation should not be a concern.

The three Fast I/O system services are:

• $IO_SETUP---Sets up an I/O.
• $IO_PERFORM[W]---Performs an I/O request.
• $IO_CLEANUP--Cleans up an I/O request.

A key concept behind the operation of the Fast I/O services is the file handle or fandle. A fandle is an opaque token that represents a "setup" I/O. A fandle is needed for each I/O outstanding from a process.

All possible setup, probing, and validation of arguments is performed off the mainline code path during application startup with calls to the $IO_SETUP system service. The I/O function, the AST address, the buffer object for the data buffer, and the IOSA buffer object are specified on input to $IO_SETUP service, and a fandle representing this setup is returned to the application.

To perform an I/O, the $IO_PERFORM system service is called, specifying the fandle, the channel, the data buffer address, the IOSA address, the length of the transfer, and the media address (VBN or LBN) of the transfer.

If the asynchronous version of this system service, $IO_PERFORM, is used to issue the I/O, then the application can wait for I/O completion using a $SYNCH specifying EFN$C_ENF and the appropriate IOSA. The synchronous form of the system service, $IO_PERFORMW, is used to issue an I/O and wait for it to complete. Optimum performance comes when the application uses AST completion; that is, the application does not issue an explicit wait for I/O completion.

To clean up a fandle, the fandle can be passed to the $IO_CLEANUP system service.

10.1.4.2 Modifying Existing Applications

Modifying an application to use the Fast I/O services requires a few source-code changes. For example:

1. A programmer adds code to create buffer objects for the IOSAs and data buffers.
2. The programmer changes the application to use the Fast I/O services. Not all $QIOs need to be converted. Only high-volume read/write I/O requests should be changed.
   A simple example is a "database writer" program, which writes modified pages back to the database. Suppose the writer can handle up to 16 simultaneous writes. At application startup, the programmer would add code to create 16 fandles by 16 $IO_SETUP system service calls.
3. In the main processing loop within the database writer program, the programmer replaces the $QIO calls with $IO_PERFORM calls. Each $IO_PERFORM call uses one of the 16 available fandles. While the I/O is in progress, the selected fandle is unavailable for use with other I/O requests. The database writer is probably using AST completion and recycling fandle, data buffer, and IOSA once the completion AST arrives.
   If the database writer routine cannot return until all dirty buffers are written (that is, it must wait for all I/O completions), then $IO_PERFORMW can be used. Alternatively $IO_PERFORM calls can be followed by $SYNCH system service calls passing the EFN$C_ENF argument to await I/O completions.
   The database writer will run faster and scale better because I/O requests now use less CPU time.
4. When the application exits, an $IO_CLEANUP system service call is done for each fandle returned by a prior $IO_SETUP system service call. Then the buffer objects are deleted. Image rundown performs fandle and buffer object cleanup on behalf of the application, but it is good form for the application to clean up properly.

The central point of synchronization for a given Fast I/O is its IOSA. The IOSA replaces the $QIO system service's IOSB argument. Larger than the IOSB argument, the byte count field in the IOSA is 64 bits and quadword aligned. Unlike the $QIO system service, Fast I/O services require the caller to supply an IOSA and require the IOSA to be part of a buffer object.

The IOSA context field can be used in place of the $QIO system service ASTPRM argument. The $QIO ASTPRM argument is typically used to pass a pointer back to the application on the completion AST to locate the user context needed for resuming a stalled user-thread. However, for the $IO_PERFORM system service, the ASTPRM on the completion AST is always the IOSA. Since there is no user-settable ASTPRM, an application can store a pointer to the user thread context for this I/O in the IOSA context field and retrieve the pointer from the IOSA in the completion AST.

10.1.4.4 $IO_SETUP

The $IO_SETUP system service performs the setup of an I/O and returns a unique identifier for this setup I/O, called a fandle, to be used on future I/Os. The $IO_SETUP arguments used to create a given fandle remain fixed throughout the life of the fandle. This has implications for the number of fandles needed in an application. For example, a single fandle can be used only for reads or only for writes. If an application module has up to 16 simultaneous reads or writes pending, then potentially 32 fandles are needed to avoid any $IO_SETUP calls during mainline processing.

The $IO_SETUP system service supports an expedite flag, which is available to boost the priority of an I/O among the other I/O requests that have been handed off to the controller. Unrestrained use of this argument is useless, because if all I/O is expedited, nothing is expedited. Note that this flag requires the use of ALTPRI and PHY_IO privilege.

10.1.4.5 $IO_PERFORM[W]

The $IO_PERFORM[W] system service accepts a fandle and five other variable I/O parameters for the high-performance I/O operation. The fandle remains in use to the application until the $IO_PERFORMW returns or if $IO_PERFORM is used until a completion AST arrives.

The CHAN argument to the fandle contains the data channel returned to the application by a previous file operation. This argument allows the application the flexibility of using the same fandle for different open files on successive I/Os. However, if the fandle is used repeatedly for the same file or channel, then an internal optimization with $IO_PERFORM is taken.

Note that $IO_PERFORM was designed to have no more than six arguments to take advantage of the OpenMS Calling Standard, which specifies that calls with up to six arguments can be passed entirely in registers.

10.1.4.6 $IO_CLEANUP

A fandle can be cleaned up by passing the fandle to the $IO_CLEANUP system service.

10.1.4.7 Fast I/O FDT Routine (ACP_STD$FASTIO_BLOCK)

Because $IO_PERFORM supports only four function codes, this system service does not use the generalized function decision table (FDT) dispatching that is contained in the $QIO system service. Instead, $IO_PERFORM uses a single vector in the driver dispatch table called DDT$PS_FAST_FDT for all the four supported functions. The DDT$PS_FAST_FDT field is a FDT routine vector that indicates whether the device driver called by $IO_PERFORM is set up to handle Fast I/O operations. A nonzero value for this field indicates that the device driver supports Fast I/O operations and that the I/O can be fully optimized.

If the DDT$PS_FAST_FDT field is zero, then the driver is not set up to handle Fast I/O operations. The $IO_PERFORM system service tolerates such device drivers, but the I/O is only slightly optimized in this circumstance.

The OpenVMS disk and tape drivers that ship as part of OpenVMS Version 7.0 have added the following line to their driver dispatch table (DDTAB) macro:

FAST_FDT=ACP_STD$FASTIO_BLOCK,- ; Fast-IO FDT routine

This line initializes the DDT$PS_FAST_FDT field to the address of the standard Fast I/O FDT routine, ACP_STD$FASTIO_BLOCK.

If you have a disk or tape device driver that can handle Fast I/O operations, you can add this DDTAB macro line to your driver. If you cannot use the standard Fast I/O FDT routine, ACP_STD$FASTIO_BLOCK, you can develop your own based on the model presented in this routine.

10.1.5 Additional Information

For complete information about the following Fast I/O system services, refer to the OpenVMS System Services Reference Manual: A--GETMSG and OpenVMS System Services Reference Manual: GETQUI--Z.

• $CREATE_BUFOBJ
• $DELETE_BUFOBJ
• $CREATE_BUFOBJ_64
• $IO_SETUP
• $IO_PERFORM
• $IO_CLEANUP

To see a sample program that demonstrates the use of buffer objects and the Fast I/O system services, refer to the IO_PERFORM.C program in the SYS$EXAMPLES directory.

10.2 Fast Path

Fast Path is an optional, high-performance feature designed to improve I/O performance. Fast Path creates a streamlined path to the device. Fast Path is of interest to any application where enhanced I/O performance is desirable. Two examples are database systems and real-time applications, where the speed of transferring data to disk is often a vital concern.

Using Fast Path features does not require source-code changes. Minor interface changes are available for expert programmers who want to maximize Fast Path benefits.

Beginning with OpenVMS Alpha Version 7.1, Fast Path supports disk I/O for the CIXCD and the CIPCA ports. These ports provide access to CI storage for XMI- and PCI-based systems. In Version 7.0, Fast Path supported disk I/O for the CIXCD port only.

Fast Path is not available on the OpenVMS VAX operating system.

10.2.1 Fast Path Features and Benefits

Fast Path achieves dramatic performance gains by reducing CPU time for I/O requests on both uniprocessor and SMP systems. These savings are on the order of 25% less CPU cost per I/O request on a uniprocessor and 35% less on a multiprocessor system. The performance benefits are produced by:

• Reducing code paths through streamlining for the case of high-volume I/O
• Substituting port-specific spinlocks for global I/O subsystem spinlocks
• Executing I/O requests for a given port on a specific CPU

The performance improvement can best be seen by contrasting the current OpenVMS I/O scheme to the new Fast Path scheme. While transparent to an OpenVMS user, each disk and tape device is tied to a specific port. All I/O for a device is sent out over its assigned port. Under the current OpenVMS I/O scheme, an I/O can be initiated on any CPU, but I/O completion must occur on the primary CPU. Under Fast Path, all I/O for a given port is assigned to a specific CPU, eliminating the requirement for completing the I/O on the primary CPU. This means that the entire I/O can be initiated and completed on a single CPU. Because I/O operations are no longer split among different CPUs, performance increases as memory cache thrashing between CPUs decreases.

Fast Path also removes the primary CPU as a possible SMP bottleneck. Without Fast Path, the primary CPU must be involved in all I/O. Once this CPU becomes saturated, no further increase in I/O throughput is possible. Spreading the I/O load evenly among CPUs in a multiprocessor system provides greater maximum I/O throughput. This is achieved by assigning each Fast Path port to a specific CPU referred to as the port's preferred CPU.

With most of the I/O code path executing under port-specific spinlocks and on each port's preferred CPU, a highly scalable SMP model of parallel I/O operation exists. Given multiple ports and CPUs, I/Os can be issued and processed in parallel to a large degree.

Preferred CPU Selection

All Fast Path ports are assignable to CPUs. You can set a SYSGEN parameter specifying the set of CPUs that are allowed to serve as preferred CPUs. This set is called the set of allowable CPUs. At any point in time, the set of CPUs that currently can have ports assigned to them, called the set of usable CPUs, is the intersection of the set of allowable CPUS, and the current set of running CPUs.

Each Fast Path Port is initially assigned to a CPU by the FASTPATH_SERVER process that runs at port initialization time. This process executes an automatic assignment algorithm that spreads Fast Path ports evenly among the usable CPUs. The FASTPATH_SERVER process also runs whenever a secondary CPU is started, and whenever the set of SYSGEN parameters specifying the allowable CPUs is changed.

If the primary CPU is in the set of allowable CPUs, the initial distribution will be biased against the primary CPU in that a port will only be assigned to the primary after ports have been assigned to each of the other usable CPUs.

To identify a device or port's current preferred CPU, you can use either $GETDVI or the SHOW DEVICE/FULL command. To identify the Fast Path ports currently assigned to a CPU, you use the SHOW CPU /FULL command.

You can directly assign a Fast Path port to a CPU, or request the system to automatically select the port's preferred CPU from a specific set of CPUs. To do this, you either issue a $QIO or use the SET DEVICE/PREFERRED_CPU command. This will also set the port's User Preferred CPU to be the selected CPU.

You can clear the port's User Preferred CPU by issuing either a $QIO, or by using the SET DEVICE/NOPREFERRED CPU DCL command.

You can redistribute the system assignable Fast Path ports across a subset of the set of usable CPUs by calling the $IO_FASTPATH system service.

Optimizing Application Performance

Processes running on a port's preferred CPU have an inherent advantage when issuing I/O to a port in that the overhead to assign the I/O to the preferred CPU can be avoided. An application process can use the $PROCESS_AFFINITY system service to assign itself to the preferred CPU of the device to which the majority of its I/O is sent.

With proper attention to assignment, a process's execution need never leave the preferred CPU. This presents a scalable process and I/O scheme for maximizing multiprocessor system operation. Like most RISC systems, Alpha system performance is highly dependent on the performance of CPU memory caches. Process assignment and preferred CPU assignment are two keys to minimizing the memory stalls in the application and in the operating system, thereby maximizing multiprocessor system throughput.

	privacy and legal statement
6136PRO_035.HTML