Ultimately, the proof of the effectiveness of the algorithms described in the previous section is the long term performance of the new file system.

Our empirical studies have shown that the inode layout policy has been effective. When running the ``list directory'' command on a large directory that itself contains many directories (to force the system to access inodes in multiple cylinder groups), the number of disk accesses for inodes is cut by a factor of two. The improvements are even more dramatic for large directories containing only files, disk accesses for inodes being cut by a factor of eight. This is most encouraging for programs such as spooling daemons that access many small files, since these programs tend to flood the disk request queue on the old file system.

Table 2 summarizes the measured throughput of the new file system. Several comments need to be made about the conditions under which these tests were run. The test programs measure the rate at which user programs can transfer data to or from a file without performing any processing on it. These programs must read and write enough data to insure that buffering in the operating system does not affect the results. They are also run at least three times in succession; the first to get the system into a known state and the second two to insure that the experiment has stabilized and is repeatable. The tests used and their results are discussed in detail in [Kridle83]\(dg. .FS \(dg A UNIX command that is similar to the reading test that we used is ``cp file /dev/null'', where ``file'' is eight megabytes long. .FE The systems were running multi-user but were otherwise quiescent. There was no contention for either the CPU or the disk arm. The only difference between the UNIBUS and MASSBUS tests was the controller. All tests used an AMPEX Capricorn 330 megabyte Winchester disk. As Table 2 shows, all file system test runs were on a VAX 11/750. All file systems had been in production use for at least a month before being measured. The same number of system calls were performed in all tests; the basic system call overhead was a negligible portion of the total running time of the tests. .KF B

1 Table 2a - Reading rates of the old and new UNIX file systems.

1 Table 2b - Writing rates of the old and new UNIX file systems. .KE

Unlike the old file system, the transfer rates for the new file system do not appear to change over time. The throughput rate is tied much more strongly to the amount of free space that is maintained. The measurements in Table 2 were based on a file system with a 10% free space reserve. Synthetic work loads suggest that throughput deteriorates to about half the rates given in Table 2 when the file systems are full.

The percentage of bandwidth given in Table 2 is a measure of the effective utilization of the disk by the file system. An upper bound on the transfer rate from the disk is calculated by multiplying the number of bytes on a track by the number of revolutions of the disk per second. The bandwidth is calculated by comparing the data rates the file system is able to achieve as a percentage of this rate. Using this metric, the old file system is only able to use about 3-5% of the disk bandwidth, while the new file system uses up to 47% of the bandwidth.

Both reads and writes are faster in the new system than in the old system. The biggest factor in this speedup is because of the larger block size used by the new file system. The overhead of allocating blocks in the new system is greater than the overhead of allocating blocks in the old system, however fewer blocks need to be allocated in the new system because they are bigger. The net effect is that the cost per byte allocated is about the same for both systems.

In the new file system, the reading rate is always at least as fast as the writing rate. This is to be expected since the kernel must do more work when allocating blocks than when simply reading them. Note that the write rates are about the same as the read rates in the 8192 byte block file system; the write rates are slower than the read rates in the 4096 byte block file system. The slower write rates occur because the kernel has to do twice as many disk allocations per second, making the processor unable to keep up with the disk transfer rate.

In contrast the old file system is about 50% faster at writing files than reading them. This is because the write system call is asynchronous and the kernel can generate disk transfer requests much faster than they can be serviced, hence disk transfers queue up in the disk buffer cache. Because the disk buffer cache is sorted by minimum seek distance, the average seek between the scheduled disk writes is much less than it would be if the data blocks were written out in the random disk order in which they are generated. However when the file is read, the read system call is processed synchronously so the disk blocks must be retrieved from the disk in the non-optimal seek order in which they are requested. This forces the disk scheduler to do long seeks resulting in a lower throughput rate.

In the new system the blocks of a file are more optimally ordered on the disk. Even though reads are still synchronous, the requests are presented to the disk in a much better order. Even though the writes are still asynchronous, they are already presented to the disk in minimum seek order so there is no gain to be had by reordering them. Hence the disk seek latencies that limited the old file system have little effect in the new file system. The cost of allocation is the factor in the new system that causes writes to be slower than reads.

The performance of the new file system is currently limited by memory to memory copy operations required to move data from disk buffers in the system's address space to data buffers in the user's address space. These copy operations account for about 40% of the time spent performing an input/output operation. If the buffers in both address spaces were properly aligned, this transfer could be performed without copying by using the VAX virtual memory management hardware. This would be especially desirable when transferring large amounts of data. We did not implement this because it would change the user interface to the file system in two major ways: user programs would be required to allocate buffers on page boundaries, and data would disappear from buffers after being written.

Greater disk throughput could be achieved by rewriting the disk drivers to chain together kernel buffers. This would allow contiguous disk blocks to be read in a single disk transaction. Many disks used with UNIX systems contain either 32 or 48 512 byte sectors per track. Each track holds exactly two or three 8192 byte file system blocks, or four or six 4096 byte file system blocks. The inability to use contiguous disk blocks effectively limits the performance on these disks to less than 50% of the available bandwidth. If the next block for a file cannot be laid out contiguously, then the minimum spacing to the next allocatable block on any platter is between a sixth and a half a revolution. The implication of this is that the best possible layout without contiguous blocks uses only half of the bandwidth of any given track. If each track contains an odd number of sectors, then it is possible to resolve the rotational delay to any number of sectors by finding a block that begins at the desired rotational position on another track. The reason that block chaining has not been implemented is because it would require rewriting all the disk drivers in the system, and the current throughput rates are already limited by the speed of the available processors.

Currently only one block is allocated to a file at a time. A technique used by the DEMOS file system when it finds that a file is growing rapidly, is to preallocate several blocks at once, releasing them when the file is closed if they remain unused. By batching up allocations, the system can reduce the overhead of allocating at each write, and it can cut down on the number of disk writes needed to keep the block pointers on the disk synchronized with the block allocation [Powell79]. This technique was not included because block allocation currently accounts for less than 10% of the time spent in a write system call and, once again, the current throughput rates are already limited by the speed of the available processors.