Boost application performance using asynchronous I/O

http://www.ibm.com/developerworks/linux/library/l-async/

Summary:  The most common input/output (I/O) model used in Linux® is synchronous I/O. After a request is made in this model, the application blocks until the request is satisfied. This is a great paradigm because the calling application requires no central processing unit (CPU) while it awaits the completion of the I/O request. But in some cases there's a need to overlap an I/O request with other processing. The Portable Operating System Interface (POSIX) asynchronous I/O (AIO) application program interface (API) provides this capability. In this article, get an overview of the API and see how to use it.

Introduction to AIO 

Linux asynchronous I/O is a relatively recent addition to the Linux kernel. It's a standard feature of the 2.6 kernel, but you can find patches for 2.4. The basic idea behind AIO is to allow a process to initiate a number of I/O operations without having to block or wait for any to complete. At some later time, or after being notified of I/O completion, the process can retrieve the results of the I/O.

I/O models 

Before digging into the AIO API, let's explore the different I/O models that are available under Linux. This isn't intended as an exhaustive review, but rather aims to cover the most common models to illustrate their differences from asynchronous I/O. Figure 1 shows synchronous and asynchronous models, as well as blocking and non-blocking models.

Figure 1. Simplified matrix of basic Linux I/O models

Each of these I/O models has usage patterns that are advantageous for particular applications. This section briefly explores each one.

I/O-bound versus CPU-bound processes

A process that is I/O bound is one that performs more I/O than processing. A CPU-bound process does more processing than I/O. The Linux 2.6 scheduler actually favors I/O-bound processes because they commonly initiate an I/O and then block, which means other work can be efficiently interlaced between them.

Synchronous blocking I/O 

One of the most common models is the synchronous blocking I/O model. In this model, the user-space application performs a system call that results in the application blocking. This means that the application blocks until the system call is complete (data transferred or error). The calling application is in a state where it consumes no CPU and simply awaits the response, so it is efficient from a processing perspective.

Figure 2 illustrates the traditional blocking I/O model, which is also the most common model used in applications today. Its behaviors are well understood, and its usage is efficient for typical applications. When theread system call is invoked, the application blocks and the context switches to the kernel. The read is then initiated, and when the response returns (from the device from which you're reading), the data is moved to the user-space buffer. Then the application is unblocked (and theread call returns).

Figure 2. Typical flow of the synchronous blocking I/O model

From the application's perspective, the read call spans a long duration. But, in fact, the application is actually blocked while the read is multiplexed with other work in the kernel.

Synchronous non-blocking I/O 

A less efficient variant of synchronous blocking is synchronous non-blocking I/O. In this model, a device is opened as non-blocking. This means that instead of completing an I/O immediately, aread may return an error code indicating that the command could not be immediately satisfied (EAGAIN orEWOULDBLOCK), as shown in Figure 3.

Figure 3. Typical flow of the synchronous non-blocking I/O model

The implication of non-blocking is that an I/O command may not be satisfied immediately, requiring that the application make numerous calls to await completion. This can be extremely inefficient because in many cases the application must busy-wait until the data is available or attempt to do other work while the command is performed in the kernel. As also shown in Figure 3, this method can introduce latency in the I/O because any gap between the data becoming available in the kernel and the user callingread to return it can reduce the overall data throughput.

Asynchronous blocking I/O 

Another blocking paradigm is non-blocking I/O with blocking notifications. In this model, non-blocking I/O is configured, and then the blockingselect system call is used to determine when there's any activity for an I/O descriptor. What makes theselect call interesting is that it can be used to provide notification for not just one descriptor, but many. For each descriptor, you can request notification of the descriptor's ability to write data, availability of read data, and also whether an error has occurred.

Figure 4. Typical flow of the asynchronous blocking I/O model (select)

The primary issue with the select call is that it's not very efficient. While it's a convenient model for asynchronous notification, its use for high-performance I/O is not advised

Asynchronous non-blocking I/O (AIO) 

Finally, the asynchronous non-blocking I/O model is one of overlapping processing with I/O. The read request returns immediately, indicating that theread was successfully initiated. The application can then perform other processing while the background read operation completes. When theread response arrives, a signal or a thread-based callback can be generated to complete the I/O transaction.

Figure 5. Typical flow of the asynchronous non-blocking I/O model

The ability to overlap computation and I/O processing in a single process for potentially multiple I/O requests exploits the gap between processing speed and I/O speed. While one or more slow I/O requests are pending, the CPU can perform other tasks or, more commonly, operate on already completed I/Os while other I/Os are initiated.

The next section examines this model further, explores the API, and then demonstrates a number of the commands.

Motivation for asynchronous I/O 

From the previous taxonomy of I/O models, you can see the motivation for AIO. The blocking models require the initiating application to block when the I/O has started. This means that it isn't possible to overlap processing and I/O at the same time. The synchronous non-blocking model allows overlap of processing and I/O, but it requires that the application check the status of the I/O on a recurring basis. This leaves asynchronous non-blocking I/O, which permits overlap of processing and I/O, including notification of I/O completion.

The functionality provided by the select function (asynchronous blocking I/O) is similar to AIO, except that it still blocks. However, it blocks on notifications instead of the I/O call.

Introduction to AIO for Linux 

This section explores the asynchronous I/O model for Linux to help you understand how to apply it in your applications.

In a traditional I/O model, there is an I/O channel that is identified by a unique handle. In UNIX®, these are file descriptors (which are the same for files, pipes, sockets, and so on). In blocking I/O, you initiate a transfer and the system call returns when it's complete or an error has occurred.
.
In asynchronous non-blocking I/O, you have the ability to initiate multiple transfers at the same time. This requires a unique context for each transfer so you can identify it when it completes. In AIO, this is an aiocb (AIO I/O Control Block) structure. This structure contains all of the information about a transfer, including a user buffer for data. When notification for an I/O occurs (called a completion), the aiocb structure is provided to uniquely identify the completed I/O. The API demonstration shows how to do this.

AIO for Linux

AIO first entered the Linux kernel in 2.5 and is now a standard feature of 2.6 production kernels.

AIO API

The AIO interface API is quite simple, but it provides the necessary functions for data transfer with a couple of different notification models. Table 1 shows the AIO interface functions, which are further explained later in this section.

Table 1. AIO interface APIs
API functionDescriptionaio_readRequest an asynchronous read operationaio_errorCheck the status of an asynchronous requestaio_returnGet the return status of a completed asynchronous requestaio_writeRequest an asynchronous operationaio_suspendSuspend the calling process until one or more asynchronous requests have completed (or failed)aio_cancelCancel an asynchronous I/O requestlio_listioInitiate a list of I/O operations

Each of these API functions uses the aiocb structure for initiating or checking. This structure has a number of elements, but Listing 1 shows only the ones that you'll need to (or can) use.

Listing 1. The aiocb structure showing the relevant fields

struct aiocb {  int aio_fildes;               // File Descriptor  int aio_lio_opcode;           // Valid only for lio_listio (r/w/nop)  volatile void *aio_buf;       // Data Buffer  size_t aio_nbytes;            // Number of Bytes in Data Buffer  struct sigevent aio_sigevent; // Notification Structure  /* Internal fields */  ...};

The sigevent structure tells AIO what to do when the I/O completes. You'll explore this structure in the AIO demonstration. Now I'll show you how the individual API functions for AIO work and how you can use them.

aio_read

The aio_read function requests an asynchronous read operation for a valid file descriptor. The file descriptor can represent a file, a socket, or even a pipe. Theaio_read function has the following prototype:

int aio_read( struct aiocb *aiocbp );

The aio_read function returns immediately after the request has been queued. The return value is zero on success or -1 on error, whereerrno is defined.

To perform a read, the application must initialize the aiocb structure. The following short example illustrates filling in theaiocb request structure and usingaio_read to perform an asynchronous read request (ignore notification for now). It also shows use of theaio_error function, but I'll explain that later.

Listing 2. Sample code for an asynchronous read with aio_read

#include <aio.h>...  int fd, ret;  struct aiocb my_aiocb;  fd = open( "file.txt", O_RDONLY );  if (fd < 0) perror("open");  /* Zero out the aiocb structure (recommended) */  bzero( (char *)&my_aiocb, sizeof(struct aiocb) );  /* Allocate a data buffer for the aiocb request */  my_aiocb.aio_buf = malloc(BUFSIZE+1);  if (!my_aiocb.aio_buf) perror("malloc");  /* Initialize the necessary fields in the aiocb */  my_aiocb.aio_fildes = fd;  my_aiocb.aio_nbytes = BUFSIZE;  my_aiocb.aio_offset = 0;  ret = aio_read( &my_aiocb );  if (ret < 0) perror("aio_read");  while ( aio_error( &my_aiocb ) == EINPROGRESS ) ;  if ((ret = aio_return( &my_iocb )) > 0) {    /* got ret bytes on the read */  } else {    /* read failed, consult errno */  }

In Listing 2, after the file from which you're reading data is opened, you zero out youraiocb structure, and then allocate a data buffer. The reference to the data buffer is placed intoaio_buf. Subsequently, you initialize the size of the buffer into aio_nbytes. The aio_offset is set to zero (the first offset in the file). You set the file descriptor from which you're reading intoaio_fildes. After these fields are set, you callaio_read to request the read. You can then make a call toaio_error to determine the status of theaio_read. As long as the status isEINPROGRESS, you busy-wait until the status changes. At this point, your request has either succeeded or failed.

Building with the AIO interface

You can find the function prototypes and other necessary symbolics in the aio.h header file. When building an application that uses this interface, you must use the POSIX real-time extensions library (librt).

Note the similarities to reading from the file with the standard library functions. In addition to the asynchronous nature ofaio_read, another difference is setting the offset for the read. In a typicalread call, the offset is maintained for you in the file descriptor context. For each read, the offset is updated so that subsequent reads address the next block of data. This isn't possible with asynchronous I/O because you can perform many read requests simultaneously, so you must specify the offset for each particular read request.

aio_error

The aio_error function is used to determine the status of a request. Its prototype is:

int aio_error( struct aiocb *aiocbp );

This function can return the following:

• EINPROGRESS, indicating the request has not yet completed
• ECANCELLED, indicating the request was cancelled by the application
• -1, indicating that an error occurred for which you can consult errno

aio_return

Another difference between asynchronous I/O and standard blocking I/O is that you don't have immediate access to the return status of your function because you're not blocking on theread call. In a standardread call, the return status is provided upon return of the function. With asynchronous I/O, you use theaio_return function. This function has the following prototype:

ssize_t aio_return( struct aiocb *aiocbp );

This function is called only after the aio_error call has determined that your request has completed (either successfully or in error). The return value ofaio_return is identical to that of theread or write system call in a synchronous context (number of bytes transferred or-1 for error).

aio_write

The aio_write function is used to request an asynchronous write. Its function prototype is:

int aio_write( struct aiocb *aiocbp );

The aio_write function returns immediately, indicating that the request has been enqueued (with a return of0 on success and-1 on failure, with errno properly set).

This is similar to the read system call, but one behavior difference is worth noting. Recall that the offset to be used is important with theread call. However, withwrite, the offset is important only if used in a file context where theO_APPEND option is not set. IfO_APPEND is set, then the offset is ignored and the data is appended to the end of the file. Otherwise, theaio_offset field determines the offset at which the data is written to the file.

aio_suspend

You can use the aio_suspend function to suspend (or block) the calling process until an asynchronous I/O request has completed, a signal is raised, or an optional timeout occurs. The caller provides a list ofaiocb references for which the completion of at least one will causeaio_suspend to return. The function prototype foraio_suspend is:

int aio_suspend( const struct aiocb *const cblist[],                  int n, const struct timespec *timeout );

Using aio_suspend is quite simple. A list of aiocb references is provided. If any of them complete, the call returns with0. Otherwise,-1 is returned, indicating an error occurred. See Listing 3.

Listing 3. Using the aio_suspend function to block on asynchronous I/Os

struct aioct *cblist[MAX_LIST]/* Clear the list. */bzero( (char *)cblist, sizeof(cblist) );/* Load one or more references into the list */cblist[0] = &my_aiocb;ret = aio_read( &my_aiocb );ret = aio_suspend( cblist, MAX_LIST, NULL );

Note that the second argument of aio_suspend is the number of elements incblist, not the number ofaiocb references. Any NULL element in thecblist is ignored byaio_suspend.

If a timeout is provided to aio_suspend and the timeout occurs, then-1is returned anderrno contains EAGAIN.

aio_cancel

The aio_cancel function allows you to cancel one or all outstanding I/O requests for a given file descriptor. Its prototype is:

int aio_cancel( int fd, struct aiocb *aiocbp );

To cancel a single request, provide the file descriptor and the aiocb reference. If the request is successfully cancelled, the function returnsAIO_CANCELED. If the request completes, the function returnsAIO_NOTCANCELED.

To cancel all requests for a given file descriptor, provide that file descriptor and aNULL reference foraiocbp. The function returns AIO_CANCELED if all requests are canceled,AIO_NOT_CANCELED if at least one request couldn't be canceled, andAIO_ALLDONE if none of the requests could be canceled. You can then evaluate each individual AIO request usingaio_error. If the request was canceled,aio_error returns-1, and errno is set to ECANCELED.

lio_listio

Finally, AIO provides a way to initiate multiple transfers at the same time using thelio_listio API function. This function is important because it means you can start lots of I/Os in the context of a single system call (meaning one kernel context switch). This is great from a performance perspective, so it's worth exploring. Thelio_listio API function has the following prototype:

int lio_listio( int mode, struct aiocb *list[], int nent,                   struct sigevent *sig );

The mode argument can be LIO_WAIT or LIO_NOWAIT.LIO_WAIT blocks the call until all I/O has completed.LIO_NOWAIT returns after the operations have been queued. Thelist is a list ofaiocb references, with the maximum number of elements defined bynent. Note that elements oflist may be NULL, whichlio_listio ignores. Thesigevent reference defines the method for signal notification when all I/O is complete.

The request for lio_listio is slightly different than the typicalread orwrite request in that the operation must be specified. This is illustrated in Listing 4.

Listing 4. Using the lio_listio function to initiate a list of requests

struct aiocb aiocb1, aiocb2;struct aiocb *list[MAX_LIST];.../* Prepare the first aiocb */aiocb1.aio_fildes = fd;aiocb1.aio_buf = malloc( BUFSIZE+1 );aiocb1.aio_nbytes = BUFSIZE;aiocb1.aio_offset = next_offset;aiocb1.aio_lio_opcode = LIO_READ;...bzero( (char *)list, sizeof(list) );list[0] = &aiocb1;list[1] = &aiocb2;ret = lio_listio( LIO_WAIT, list, MAX_LIST, NULL );

The read operation is noted in the aio_lio_opcode field with LIO_READ. For a write operation, LIO_WRITE is used, but LIO_NOP is also valid for no operation.

AIO notifications

Now that you've seen the AIO functions that are available, this section digs into the methods that you can use for asynchronous notification. I'll explore asynchronous notification through signals and function callbacks.

Asynchronous notification with signals

The use of signals for interprocess communication (IPC) is a traditional mechanism in UNIX and is also supported by AIO. In this paradigm, the application defines a signal handler that is invoked when a specified signal occurs. The application then specifies that an asynchronous request will raise a signal when the request has completed. As part of the signal context, the particularaiocb request is provided to keep track of multiple potentially outstanding requests. Listing 5 demonstrates this notification method.

Listing 5. Using signals as notification for AIO requests

void setup_io( ... ){  int fd;  struct sigaction sig_act;  struct aiocb my_aiocb;  ...  /* Set up the signal handler */  sigemptyset(&sig_act.sa_mask);  sig_act.sa_flags = SA_SIGINFO;  sig_act.sa_sigaction = aio_completion_handler;  /* Set up the AIO request */  bzero( (char *)&my_aiocb, sizeof(struct aiocb) );  my_aiocb.aio_fildes = fd;  my_aiocb.aio_buf = malloc(BUF_SIZE+1);  my_aiocb.aio_nbytes = BUF_SIZE;  my_aiocb.aio_offset = next_offset;  /* Link the AIO request with the Signal Handler */  my_aiocb.aio_sigevent.sigev_notify = SIGEV_SIGNAL;  my_aiocb.aio_sigevent.sigev_signo = SIGIO;  my_aiocb.aio_sigevent.sigev_value.sival_ptr = &my_aiocb;  /* Map the Signal to the Signal Handler */  ret = sigaction( SIGIO, &sig_act, NULL );  ...  ret = aio_read( &my_aiocb );}void aio_completion_handler( int signo, siginfo_t *info, void *context ){  struct aiocb *req;  /* Ensure it's our signal */  if (info->si_signo == SIGIO) {    req = (struct aiocb *)info->si_value.sival_ptr;    /* Did the request complete? */    if (aio_error( req ) == 0) {      /* Request completed successfully, get the return status */      ret = aio_return( req );    }  }  return;}

In Listing 5, you set up your signal handler to catch the SIGIO signal in theaio_completion_handler function. You then initialize theaio_sigevent structure to raiseSIGIO for notification (which is specified via theSIGEV_SIGNAL definition insigev_notify). When your read completes, your signal handler extracts the particularaiocb from the signal'ssi_value structure and checks the error status and return status to determine I/O completion.

For performance, the completion handler is an ideal spot to continue the I/O by requesting the next asynchronous transfer. In this way, when completion of one transfer has completed, you immediately start the next.

Asynchronous notification with callbacks

An alternative notification mechanism is the system callback. Instead of raising a signal for notification, this mechanism calls a function in user-space for notification. You initialize theaiocb reference into thesigevent structure to uniquely identify the particular request being completed; see Listing 6.

Listing 6. Using thread callback notification for AIO requests

void setup_io( ... ){  int fd;  struct aiocb my_aiocb;  ...  /* Set up the AIO request */  bzero( (char *)&my_aiocb, sizeof(struct aiocb) );  my_aiocb.aio_fildes = fd;  my_aiocb.aio_buf = malloc(BUF_SIZE+1);  my_aiocb.aio_nbytes = BUF_SIZE;  my_aiocb.aio_offset = next_offset;  /* Link the AIO request with a thread callback */  my_aiocb.aio_sigevent.sigev_notify = SIGEV_THREAD;  my_aiocb.aio_sigevent.notify_function = aio_completion_handler;  my_aiocb.aio_sigevent.notify_attributes = NULL;  my_aiocb.aio_sigevent.sigev_value.sival_ptr = &my_aiocb;  ...  ret = aio_read( &my_aiocb );}void aio_completion_handler( sigval_t sigval ){  struct aiocb *req;  req = (struct aiocb *)sigval.sival_ptr;  /* Did the request complete? */  if (aio_error( req ) == 0) {    /* Request completed successfully, get the return status */    ret = aio_return( req );  }  return;}

In Listing 6, after creating your aiocb request, you request a thread callback usingSIGEV_THREAD for the notification method. You then specify the particular notification handler and load the context to be passed to the handler (in this case, a reference to theaiocb request itself). In the handler, you simply cast the incomingsigval pointer and use the AIO functions to validate the completion of the request.

System tuning for AIO

The proc file system contains two virtual files that can be tuned for asynchronous I/O performance:

• The /proc/sys/fs/aio-nr file provides the current number of system-wide asynchronous I/O requests.
• The /proc/sys/fs/aio-max-nr file is the maximum number of allowable concurrent requests. The maximum is commonly 64KB, which is adequate for most applications.

Summary

Using asynchronous I/O can help you build faster and more efficient I/O applications. If your application can overlap processing and I/O, then AIO can help you build an application that more efficiently uses the CPU resources available to you. While this I/O model differs from the traditional blocking patterns found in most Linux applications, the asynchronous notification model is conceptually simple and can simplify your design.

Resources

Learn

• The POSIX.1b implementation explains the internal details of AIO from the GNU Library perspective.
  
  
• Realtime Support in Linux explains more about AIO and a number of real-time extensions, from scheduling and POSIX I/O to POSIX threads and high resolution timers (HRT).
  
  
• In the Design Notes for the 2.5 integration, learn about the design and implementation of AIO in Linux.
  
  
• In the developerWorks Linux zone, find more resources for Linux developers.
  
  
• Stay current with developerWorks technical events and Webcasts.
  
  

Get products and technologies

• With IBM trial software, available for download directly from developerWorks, build your next development project on Linux.
  
  

Discuss

• Check out developerWorks blogs and get involved in thedeveloperWorks community.