File System Performance

File System Performance

Access to disk is much slower than access to memory. Reading a 32-bit memory word might take 10 nsec. Reading from a hard disk might proceed at 100 MB/sec, which is four times slower per 32-bit word, but to this must be added 5-10 msec to seek to the track and then wait for the desired sector to arrive under the read head. If only a single word is required, the memory access is on the order of a million times as fast as disk access. As a result of this difference in access time, many file systems have been designed with many optimizations to improve performance. In this section we will study three of them.


The most common technique used to reduce disk accesses is the block cache or buffer cache. (Cache is pronounced "cash" and is derived from the French cacher, meaning to hide.) In this context, a cache is a collection of blocks that logically belong on the disk but are being kept in memory for performance reasons.

Many algorithms can be used to manage the cache, but a common one is to check all read requests to see if the required block is in the cache. If it is, the read request can be satisfied without a disk access. If the block is not in the cache, it is first read into the cache and then copied to wherever it is required. Subsequent requests for the same block can be satisfied from the cache.

Operation of the cache is shown in Figure 1. Since there are many (often thousands of) blocks in the cache, some way is required to determine quickly if a given block is present. The usual way is to hash the device and disk address and look up the result in a hash table. All the blocks with the same hash value are chained together on a linked list so that the collision chain can be followed.

When a block has to be loaded into a full cache, some block has to be removed (and rewritten to the disk if it has been modified since being brought in). This situation is very much like paging, and all the usual page replacement algorithms explained in "MEMORY MANAGEMENT", such as FIFO, second chance, and LRU, are applicable. One pleasant difference between paging and caching is that cache references

The buffer cache data structures

are relatively infrequent, so that it is feasible to keep all the blocks in exact LRU order with linked lists.

In Figure 1, we see that in addition to the collision chains starting at the hash table, there is also a bidirectional list running through all the blocks in the order of usage, with the least recently used block on the front of this list and the most recently used block at the end of this list. When a block is referenced, it can be removed from its position on the bidirectional list and put at the end. Thus, exact LRU order can be maintained.

Unfortunately, there is a catch. Now that we have a situation in which exact LRU is possible, it turns out that LRU is undesirable. The problem has to do with the crashes and file system consistency discussed in the previous section. If a critical block, such as an i-node block, is read into the cache and modified, but not rewritten to the disk, a crash will leave the file system in an inconsistent state. If the i-node block is put at the end of the LRU chain, it may be quite a while before it reaches the front and is rewritten to the disk.

Moreover, some blocks, such as i-node blocks, are rarely referenced two times within a short interval. These considerations lead to a modified LRU scheme, taking two factors into account:

1. Is the block likely to be needed again soon?
2. Is the block essential to the consistency of the file system?

For both questions, blocks can be divided into categories such as i-node blocks, indirect blocks, directory blocks, full data blocks, and partially full data blocks. Blocks that will probably not be required again soon go on the front, rather than the rear of the LRU list, so their buffers will be reused quickly. Blocks that might be required again soon, such as a partly full block that is being written, go on the end of the list, so they will stay around for a long time.

The second question is independent of the first one. If the block is necessary to the file system consistency (basically, everything  except data blocks), and it has been modified, it should be written to disk immediately, regardless of which end of the LRU list it is put on. By writing critical blocks quickly, we greatly reduce the probability that a crash will wreck the file system. While a user may be unhappy if one of his files is ruined in a crash, he is likely to be far more unhappy if the whole file system is lost.

Even with this measure to keep the file system integrity intact, it is undesirable to keep data blocks in the cache too long before writing them out. Consider the plight of someone who is using a personal computer to write a book. Even if our writer periodically tells the editor to write the file being edited to the disk, there is a good chance that everything will still be in the cache and nothing on the disk. If the system crashes, the file system structure will not be corrupted, but a whole day's work will be lost.

This situation need not happen very often before we have a fairly unhappy user. Systems take two approaches to dealing with it. The UNIX way is to have a system call, sync, which forces all the modified blocks out onto the disk immediately. When the system is started up, a program, usually called update, is started up in the background to sit in an endless loop issuing sync calls, sleeping for 30 sec between calls. As a result, no more than 30 seconds of work is lost due to a crash.

Although Windows now has a system call equivalent to sync, FlushFileBuffers, in the past it did not. Instead, it had a different strategy that was in some ways better than the UNIX approach (and in some ways worse). What it did was to write every modified block to disk as soon as it has been written to the cache. Caches in which all modified blocks are written back to the disk immediately are called write-through caches. They require more disk I/O than nonwrite-through caches.

The difference between these two approaches can be seen when a program writes a 1-KB block full, one character at a time.  UNIX will collect all the characters in the cache and write the block out once every 30 seconds, or whenever the block is removed from the cache. With a write-through cache, there is a disk access for every character written. Of course, most programs do internal buffering, so they normally write not a character, but a line or a larger unit on each write system call.

A consequence of this difference in caching strategy is that just removing a (floppy) disk from a UNIX system without doing a sync will almost always result in lost data, and frequently in a corrupted file system as well. With write-through caching no problem arises. These differing strategies were chosen because UNIX was developed in an environment in which all disks were hard disks and not removable, whereas the first Windows file system was inherited from MS-DOS, which started out in the floppy disk world. As hard disks became the norm, the UNIX approach, with its better efficiency (but worse reliability), became the norm, and is also used now on Windows for hard disks. On the other hand, NTFS takes other measures (journaling) to improve reliability, as discussed earlier.

Some operating systems integrate the buffer cache with the page cache. This is especially attractive when memory-mapped files are supported. If a file is mapped onto memory, then some of its pages may be in memory because they were demand paged in. Such pages are hardly different from file blocks in the buffer cache. In this case, they can be treated the same way, with a single cache for both file blocks and pages.

Block Read Ahead

A second technique for improving perceived file system performance is to try to get blocks into the cache before they are required to increase the hit rate. Particularly, many files are read sequentially. When the file system is asked to produce block k in a file, it does that, but when it is finished, it makes a sneaky check in the cache to see if block k + 1 is already there. If it is not, it schedules a read for block k + 1 in the hope that when it is required, it will have already arrived in the cache. At the very least, it will be on the way.

Of course, this read ahead strategy only works for files that are being read sequentially. If a file is being randomly accessed, read ahead does not help. Actually, it hurts by tying up disk bandwidth reading in useless blocks and removing potentially useful blocks from the cache (and possibly tying up more disk bandwidth writing them back to disk if they are dirty). To see whether read ahead is worth doing, the file system can keep track of the access patterns to each open file. For instance, a bit associated with each file can keep track of whether the file is in "sequential access mode" or "random access mode". In the beginning, the file is given the benefit of the doubt and put in sequential access mode. On the other hand, whenever a seek is done, the bit is cleared. If sequential reads start happening again, the bit is set once again. Thus, the file system can make a reasonable guess about whether it should read ahead or not. If it gets it wrong once in a while, it is not a disaster, just a little bit of wasted disk bandwidth.

Reducing Disk Arm Motion

Caching and read ahead are not the only ways to increase file system performance. Another important technique is to reduce the amount of disk arm motion by putting blocks that are likely to be accessed in sequence close to each other, preferably in the same cylinder. When an output file is written, the file system has to allocate the blocks one at a time, on demand. If the free blocks are recorded in a bitmap, and the whole bitmap is in main memory, it is easy enough to choose a free block as close as possible to the previous block. With a free list, part of which is on disk, it is much harder to allocate blocks close together.

On the other hand, even with a free list, some block clustering can be done. The trick is to keep track of disk storage not in blocks, but in groups of consecutive blocks. If all sectors consist of 512 bytes, the system could use 1-KB blocks (2 sectors) but allocate disk storage in units of 2 blocks (4 sectors). This is not the same as having a 2-KB disk blocks, since the cache would still use 1-KB blocks and disk transfers would still be 1 KB, but reading a file sequentially on an otherwise idle system would reduce the number of seeks by a factor of two, considerably improving performance. A variation on the same theme is to take account of rotational positioning. When allocating blocks, the system attempts to place consecutive blocks in a file in the same cylinder.

Another performance bottleneck in systems that use i-nodes or anything like them is that reading even a short file requires two disk accesses: one for the i-node and one for the block. The usual i-node placement is shown in Figure 2(a). Here all the i-nodes are near the beginning of the disk, so the average distance between an i-node and its blocks will be about half the number of cylinders, requiring long seeks.

Inodes placed at the start of the disk

One easy performance improvement is to put the i-nodes in the middle of the disk, rather than at the start, thus reducing the average seek between the i-node and the first block by a factor of two. Another idea, shown in Figure 2(b), is to divide the disk into cylinder groups, each with its own i-nodes, blocks, and free list (McKusick et al., 1984). When creating a new file, any i-node can be chosen, but an attempt is made to find a block in the same cylinder group as the i-node. If none is available, then a block in a nearby cylinder group is used.


buffer cache, block cache, fifo, lru, page cache, i-node, write-through caches