The arrow of time

Ivan Voras' blog

Home · cp(1) and cat(1) in the Century of the Fruitbat

cp(1) and cat(1) in the Century of the Fruitbat

FreeBSD has frugal ancestors, created when memory was at premium. The code that was once optimal for the hardware, even using heuristics to adapt to various circumstances, might be underperforming now. An excellent example for this was the transition from a system highly optimized for single-processor hardware to multi-processor (SMP) hardware, which is only recently being completed. In a much simple case, cp(1) and cat(1) were using buffers of 64 kB (and sometimes even 1 kB) to copy data. This is not exactly an important aspect of the system, but I've patched it anyway :)

For different reasons, cp(1) and cat(1) were both using 64 kB buffers to copy data. Cat was trying to find out what the block size on the file system is (which, for various reasons, ends up almost always being 64 kB) and used 1 kB for non-files, and cp just gave up and used MAXBSIZE (maximum file system block size), which is predefined to be 64 kB. This is less useful today when drives and memory caches are big, and especially with file systems like ZFS where "block size" isn't even the same file system concept as in UFS.

This is not nearly as bad as it might seem because file system caching and prereading make the data available before it's needed so the effects of small reads are very minimal. A related problem here is that the kernel MAXPHYS value (maximum IO block size) is defined to be 128 kB, which puts the upper cap on IO transfers to and from disk drives. This is also not very noticable, for similar reasons, in regular usage but might be in highly loaded servers with large RAID arrays. Unfortunately this particular setting is entrenched in the kernel and disk controller drivers so it will require much restructuring. Every now and then there are signals that someone is working on increasing MAXPHYS so it might happen soon.

Returning to cp and cat, the buffer for both is now set to 128 kB (MAXPHYS) if the machine has 128 MB or less physical memory, or 1 MB (8*MAXPHYS) if there's more memory available. The most visible improvement is that the number of context switches is reduced now:

before:
> /usr/bin/time -hlp cat file1 > file2
...
       163  voluntary context switches
     11194  involuntary context switches
after:
> /usr/bin/time -hlp ./cat file1 > file2
       417  voluntary context switches
       272  involuntary context switches

There are no noticable improvements in performance, but this larger buffer size will maybe pay out better in the future, on RAM drives, or on really, really overloaded systems :)

(p.s. cp copies files smaller than 8 MB by mmap+write... this remains untouched).

Posted on 2008-10-30T19:43 by ivoras

Nice Thanks !

Added on 2008-10-30T20:21 by r3inforce

This keeps the System from aging. And its better to not rely on some magic optimizations instead of doing it right, - right ?

Post your comment here!

Your name:
Comment title:
Text:
Type "xxx" here:

Comments are subject to moderation and will be deleted if deemed inappropriate. All content is © Ivan Voras. Comments are owned by their authors (who agree to basically surrender all rights by publishing them :) )..