Here's how I/O schedulers contribute to disk performance in Linux and the improvements you can get from the new I/O schedulers in the 2.6 kernel.
Although most Linux users are familiar with the role of process schedulers, such as the new O(1) scheduler, many users are not so familiar with the role of I/O schedulers. I/O schedulers are similar in some aspects to process schedulers; for instance, both schedule some resource among multiple users. A process scheduler virtualizes the resource of processor time among multiple executing processes on the system. So, what does an I/O scheduler schedule?
A naïve system would not even include an I/O scheduler. Unlike the process scheduler, the I/O scheduler is not a mandatory component of the operating system. Instead, performance is the I/O scheduler's sole raison d'être.
To understand the role of an I/O scheduler, let's go over some background information and then look at how a system behaves without an I/O scheduler. Hard disks address their data using the familiar geometry-based addressing of cylinders, heads and sectors. A hard drive is composed of multiple platters, each consisting of a single disk, spindle and read/write head. Each platter is divided further into circular ring-like tracks, similar to a CD or record. Finally, each track is composed of some integer number of sectors.
To locate a specific unit of data in a hard drive, the drive's logic requires three pieces of information: the cylinder, the head and the sector. The cylinder specifies the track on which the data resides. If you lay the platters on top of one another (as they are in a hard disk), a given track forms a cylinder through each platter. The head then identifies the exact read/write head (and thus the exact platter) in question. The search now is narrowed down to a single track on a single platter. Finally, the sector value denotes the exact sector on the track. The search is complete: the hard disk knows what sector, on what track, on what platter the data resides. It can position the read/write head of the correct platter over the correct track and read the proper sector.
Thankfully, modern hard disks do not force computers to communicate with them in terms of cylinders, heads and sectors. Instead, modern hard drives map a unique block number over each cylinder/head/sector triplet. The unique number identifies a specific cylinder/head/sector value. Modern operating systems then can address hard drives using this block number—called logical block addressing—and the hard drive translates the block number into the correct cylinder/head/sector value.
One thing of note about this block number: although nothing guarantees it, the physical mapping tends to be sequential. That is, logical block n tends to be physically adjacent to logical block n+1. We discuss why that is important later on.
Now, let's consider the typical UNIX system. Applications as varied as databases, e-mail clients, Web servers and text editors issue I/O requests to the disk, such as read this block and write to that block. The blocks tend to be located physically all over the disk. The e-mail spool may be located in an entirely different region of the disk from the Web server's HTML data or the text editor's configuration file. Indeed, even a single file can be strewn all over the disk if the file is fragmented, that is, not laid out in sequential blocks. Because the files are broken down into individual blocks, and hard drives are addressed by block and not the much more abstract concepts of files, reading or writing file data is broken down into a stream of many individual I/O requests, each to a different block. With luck, the blocks are sequential or at least physically close together. If the blocks are not near one another, the disk head must move to another location on the disk. Moving the disk head is called seeking, and it is one of the most expensive operations in a computer. The seek time on modern hard drives is measured in the tens of milliseconds. This is one reason why defragmented files are a good thing.
Unfortunately, it does not matter if the files are defragmented because the system is generating I/O requests for multiple files, all over the disk. The e-mail client wants a little from here and the Web server wants a little from there—but wait, now the text editor wants to read a file. The net effect is that the disk head is made to jump around the disk. In a worst-case scenario, with interleaved I/O requests to multiple files, the head can spend all of its time jumping around from one location to another—not a good thing for overall system performance.
This is where the I/O scheduler comes in. The I/O scheduler schedules the pending I/O requests in order to minimize the time spent moving the disk head. This, in turn, minimizes disk seek time and maximizes hard disk throughput.
This magic is accomplished through two main actions, sorting and merging. First, the I/O scheduler keeps the list of pending I/O requests sorted by block number. When a new I/O request is issued, it is inserted, block-wise, into the list of pending requests. This prevents the drive head from seeking all around the disk to service I/O requests. Instead, by keeping the list sorted, the disk head moves in a straight line around the disk. If the hard drive is busy servicing a request at one part of the disk, and a new request comes in to the same part of the disk, that request can be serviced before moving off to other parts of the disk.
Merging occurs when an I/O request is issued to an identical or adjacent region of the disk. Instead of issuing the new request on its own, it is merged into the identical or adjacent request. This minimizes the number of outstanding requests.
Let's look at an example. Consider the case where two applications issue requests to the following block numbers, such that they arrived in the kernel in this order: 10, 500, 12, 502, 14, 504 and 12. The I/O scheduler-less approach would service these blocks in the given order. That is seven long seeks, back and forth between two parts of the disk. What a waste! If the kernel sorted and merged these requests, however, and serviced them in that order, the results would be much different: 10, 12, 14, 500, 502 and 504. Only a single far-off seek and one less request overall.
In this manner, an I/O scheduler virtualizes the resources of disk I/O among multiple I/O requests in order to maximize global throughput. Because I/O throughput is so crucial to system performance and because seek time is so horribly slow, the job of an I/O scheduler is important.
The I/O scheduler found in the 2.4 Linux kernel is named the Linus Elevator. I/O schedulers often are called elevator algorithms, because they tackle a problem similar to that of keeping an elevator moving smoothly in a large building. The Linus Elevator functions almost exactly like the classic I/O scheduler described above. For the most part, this was great because simplicity is a good thing and the 2.4 kernel's I/O scheduler just worked. Unfortunately, in the I/O scheduler's quest to maximize global I/O throughput, a trade-off was made: local fairness—in particular, request latency—can go easily out the window. Let's look at an example.
Consider a stream of requests to logical disk blocks such as 20, 30, 700 and 25. The I/O scheduler's sorting algorithm would queue and issue the requests in the following order (assuming the disk head currently is at the logical start of the disk): 20, 25, 30 and 700. This is expected and indeed correct. Assume, however, that halfway through servicing the request to block 25, another request comes in to the same part of the disk. And then another. And yet another. It is entirely feasible that the request way over to block 700 is not serviced for a relatively long time.
Worse, what if the request was to read a disk block? Read requests generally are synchronous. When an application issues a request to read some data, it typically blocks and waits until the kernel returns the data. The application must sit and wait, twiddling its thumbs, until that request way over at block 700 finally is serviced. Writes, on the other hand, typically are not synchronous—they are asynchronous. When an application issues a write, the kernel copies the data and metadata into the kernel, prepares a buffer to hold the data and returns to the application. The application does not really care or even know when the data actually hits the disk.
It gets worse for our friend the read request, however. Because writes are asynchronous, writes tend to stream. That is, it is common for a large writeback of a lot of data to occur. This implies that many individual write requests are submitted to a close area of the hard disk. As an example, consider saving a large file. The application dumps write requests on the system and hard drive as fast as it is scheduled.
Read requests, conversely, usually do not stream. Instead, applications submit read requests in small one-by-one chunks, with each chunk dependent on the last. Consider reading all of the files in a directory. The application opens the first file, issues a read request for a suitable chunk of the file, waits for the returned data, issues a read request for the next chunk, waits and continues likewise until the entire file is read. Then the file is closed, the next file is opened and the process repeats. Each subsequent request has to wait for the previous, which means substantial delays to this application if the requests are to far-off disk blocks. The phenomenon of streaming write requests starving dependent read requests is called writes-starving-reads (see Sidebar “Test 1. Writes-Starving-Reads”).
The possibility of not servicing some requests in a reasonable amount of time is called starvation. Request starvation results in unfairness. In the case of I/O schedulers, the system explicitly has decided to trade fairness for improved global throughput. That is, the system attempts to improve the overall performance of the system at the possible expense of any one specific I/O request. This is accepted and, indeed, desired—except that prolonged starvation is bad. Starving read requests for even moderate lengths of time results in high application latency for applications issuing read requests during other disk activity. This high read latency adversely affects system performance and feel (see Sidebar “Test 2. Effects of High Read Latency”).
Preventing the starvation of requests in general, and read requests in particular, was a goal of the new 2.6 I/O schedulers.
The Deadline I/O Scheduler was introduced to solve the starvation issue surrounding the 2.4 I/O scheduler and traditional elevator algorithms in general. As discussed, the Linus Elevator maintains the sorted list of pending I/O requests in a single queue. The I/O request at the head of the queue is the next one to be serviced. The Deadline I/O Scheduler continues to keep this queue, but kicks things up a notch by introducing two additional queues—the read FIFO queue and the write FIFO queue. The Deadline I/O Scheduler keeps the items in each of these queues sorted by submission time (effectively, first in is first out). The read FIFO queue, as its name suggests, contains only read requests. The write FIFO queue, likewise, contains only write requests. Each FIFO queue is assigned an expiration value. The read FIFO queue has an expiration time of 500 milliseconds. The write FIFO queue has an expiration time of five seconds.
When a new I/O request is submitted, it is insertion-sorted into the standard queue and placed at the tail of its respective (either read or write) FIFO queue. Normally, the hard drive is sent I/O requests from the head of the standard sorted queue. This maximizes global throughput by minimizing seeks, as the normal queue is sorted by block number, as with the Linus Elevator.
When the item at the head of one of the FIFO queues, however, grows older than the expiration value associated with its queue, the I/O scheduler stops dispatching I/O requests from the standard queue. Instead, it services the I/O request at the head of the FIFO queue, plus a couple extra for good measure. The I/O scheduler needs to check and handle only the requests at the head of the FIFO queues, as those are the oldest requests in the queue.
Remember our old friend, the request to block 700? Despite the flood of write requests to the faraway blocks, after 500 milliseconds the Deadline I/O Scheduler would stop servicing those requests and service the read over at block 700. The disk would seek to block 700, service the read request and then continue servicing the other outstanding requests.
In this manner, the Deadline I/O Scheduler can enforce a soft deadline on I/O requests. Although it makes no promise that an I/O request is serviced before the expiration time, the I/O scheduler generally services requests near their expiration times. Thus, the Deadline I/O Scheduler continues to provide good global throughput without starving any one request for an unacceptably long time. Because read requests are given short expiration times, the writes-starving-reads problem is minimized.
This is all well and good, but it's not a perfect solution. Consider what happens with our fictional request to block 700, which presumably is the first of many dependent reads to that area of the disk. After servicing the read request, the Deadline I/O Scheduler continues servicing the write requests to the earlier blocks. This is fine, until the reading application submits its next read request (say, to block 710). In 500 milliseconds, that request expires and the disk seeks over to block 710, services the request, seeks back to where it was before and continues servicing the streaming write. And then another read arrives.
The problem again stems from those darn dependent reads. Because reads are issued in dependent chunks, the application issues the next read only when the previous is returned. But by the time the application receives the read data, is scheduled to run and submits the next read, the I/O scheduler has moved on and begun servicing some other requests. This results in a wasted pair of seeks for each read: seek to the read, service it and seek back. If only there was some way for the I/O scheduler to know—nay, to anticipate—that another read would soon be submitted to the same part of the disk. Instead of seeking back and forth, it could wait in anticipation for the next read. Saving those awful seeks certainly is worth a few milliseconds of waiting; we save two seeks.
This is, of course, exactly what the Anticipatory I/O Scheduler does. It began as the Deadline I/O Scheduler; it implements the same deadline-based scheduling. But it was gifted with the addition of an anticipation mechanism. When a read request is submitted, the Anticipatory I/O Scheduler services it within its deadline, as usual. Unlike the Deadline I/O Scheduler, however, the Anticipatory I/O Scheduler then sits and waits, doing nothing, for up to six milliseconds. Chances are good that the application will issue another read to the same part of the filesystem during those six milliseconds. If so, that request is serviced immediately, and the Anticipatory I/O Scheduler waits some more. If six milliseconds go by without a read request, the Anticipatory I/O Scheduler guessed wrong and returns to whatever it was doing before.
If even a moderate number of requests are anticipated correctly, a great deal of time (two expensive seeks, each) is saved (Table 1). Because most reads are dependent, the anticipation pays off most of the time. To further improve the odds of a correct anticipation, the Anticipatory I/O Scheduler uses a heuristic to better guess for which processes to wait. To this end, the scheduler maintains I/O statistics about each process to keep track of its behavior. Because of these statistics and intelligent heuristics, the Anticipatory I/O Scheduler correctly anticipates the actions of applications a sufficiently large amount of the time to be well worth the overhead.
Table 1. The Results
|I/O Scheduler and Kernel||Test 1||Test 2|
|Linus Elevator on 2.4||45 seconds||30 minutes, 28 seconds|
|Deadline I/O Scheduler on 2.6||40 seconds||3 minutes, 30 seconds|
|Anticipatory I/O Scheduler on 2.6||4.6 seconds||15 seconds|
By minimizing unneeded seeks and more quickly servicing read requests, in many workloads the Anticipatory I/O Scheduler provides both improved request latency and global throughput over the Deadline I/O Scheduler and the Linus Elevator. Unsurprisingly, the Anticipatory I/O Scheduler is the default I/O scheduler in the 2.6 kernel. And rightly so, it rocks.