Computer Clocks

Since most of the examples in this text measure a time interval, we need to describe in more detail the type of timekeeping used by current Unix systems. The following description applies to the systems being used for the examples in this book, and for most Unix systems. Additional details are given in Sections 3.4 and 3.5 of [Leffler et al. 1989].

The hardware generates a clock interrupt at some frequency. For Sun SPARCs and Intel 80386s the interrupts occur every 10 ms.

It should be noted that most computers use an uncompensated crystal oscillator to generate these interrupts. As noted in Table 7 of RFC 1305 [Mills 1992], you don't want to ask what the drift per day of such an oscillator is. This means few computers keep accurate time (i.e., the interrupts don't occur exactly every 10 ms). A 0.01% tolerance gives an error of 8.64 seconds per day. To keep better time requires (1) a better oscillator, (2) an external time source with greater precision (e.g., the time source supplied by the Global Positioning Satellites), or (3) access across the Internet to systems with more precise clocks. The latter is provided by the Network Time Protocol, as described in detail in RFC 1305, which is beyond the scope of this book.

Another common source of time errors in Unix systems is that the 10-ms clock interrupts only cause the kernel to increment a variable that keeps track of the time. If the kernel loses an interrupt (i.e., it's too busy for the 10 ms between two consecutive interrupts), the clock will lose 10 ms. Lost interrupts of this type often cause Unix systems to lose time.

Even though the clock interrupts arrive approximately every 10 ms, newer systems such as SPARCs provide a higher resolution timer to measure time differences. tcpdump, through the NIT driver (described in Appendix A) has access to this higher resolution timer. On SPARCs this timer provides microsecond resolution. Access to this higher resolution timer is also provided for user processes through the gettimeofday(2) function.

The author ran the following experiment. A program was run that called the gettimeofday function 10,000 times in a loop, saving each return value in an array. At the end of the loop the 9,999 differences were printed out. For a SPARC ELC the distribution of the differences are shown in Figure B.1.


Figure B.1 Distribution of time required to call gettimeofday 10,000 times on SPARC ELC.

The total clock time required to run the program was 0.38 seconds, on an otherwise idle system. From this we can say that the time for the process to call gettimeofday is about 37 microseconds. Since the ELC is rated around 21 MIPS (million instructions per second), 37 microseconds corresponds to about 800 instructions. This seems reasonable for the kernel to handle a system call from a user process, execute the system call, copy back 8 bytes of results, and return to the user process. (MIPS ratings are questionable, and it's hard to try to measure instruction times on current systems. All we're trying to do is get a rough idea and see if the values make sense.)

From this simple experiment we can say that the values returned by gettimeofday do contain microsecond resolution.

If we run similar tests under SVR4/386, however, the results are different. This is because many 386 Unix systems, such as SVR4, only count the 10-ms clock interrupts, and don't try to provide any higher resolution. Figure B.2 is the distribution of the 9,999 differences under SVR4 on an 25 Mhz 80386.


Figure B.2 Distribution of time required to call gettimeofday 10,000 times under SVR4/386.

These values are worthless, since the differences are normally less than 10 ms, which is treated as 0. About all we can do on these systems is measure the clock time on an idle system, and divide by the number of loops. This provides an upper bound, since it includes the time required to call printf 9,999 times, writing the results to a file. (In the SPARC case. Figure B.1, the differences did not include the printf times since all 10,000 values were first obtained, and then the results were printed.) Under SVR4 the clock time was 3.15 seconds, yielding 315 microseconds per system call. This system call time, about 8.5 times slower than the SPARC, seems about right.

BSD/386 Version 1.0 provides microsecond resolution similar to the SPARC. It reads the 8253 clock register and calculates the number of microseconds since the last clock tick. This is made available to processes that call gettimeofday and to kernel modules such as the BSD Packet Filter.

In relation to tcpdump these numbers mean that we can believe the millisecond and submillisecond values that are printed on the SPARC and BSD/386 systems, but the values printed by tcpdump under SVR4/386 will always be a multiple of 10 ms. For other programs that print round-trip times, such as ping (Chapter 7) and traceroute (Chapter 8), on the SPARC and BSD/386 systems we can believe the millisecond values that are output, but the values printed under SVR4/386 will always be multiples of 10. To measure anything like the ping time on a LAN, which we show in Chapter 7 to be around 3 ms, requires running ping on the SPARC or BSD/386.

Some of the examples in this text were run under BSD/386 Version 0.9.4, which was similar to SVR4 in that it provided only 10-ms clock resolution. When we show tcpdump output from this system, we show only two numbers to the right of the decimal point, since that's the resolution provided.