misc/34596: slow gettimeofday in FreeBSD 4.5

[Bruce Evans]

The following reply was made to PR misc/34596; it has been noted by GNATS.

From: Bruce Evans <bde at zeta.org.au>
To: Andreas Gustafsson <gson at nominum.com>
Cc: freebsd-gnats-submit at freebsd.org
Subject: Re: misc/34596: slow gettimeofday in FreeBSD 4.5
Date: Wed, 4 Jun 2003 20:26:10 +1000 (EST)

 >  gettimeofday() is still really slow in FreeBSD 4.8-STABLE.  Some
 >  numbers from the test program submitted with misc/34596:
 >
 >    FreeBSD 4.8, single Intel Xeon 2.8 GHz          delta: 5s 59701us
 >    NetBSD 1.6L, dual AMD Athlon MP 1800+           delta: 2s 801360us
 >    NetBSD 1.6T, single Transmeta Crusoe 600 MHz    delta: 2s 582755us
 >    Linux 2.4.9-31smp, dual AMD Athlon MP 1900+     delta: 0s 247116us
 >
 >  That's more than 5 microseconds on a 2.8 GHz processor, or more than
 >  14000 clock cycles per call.
 >
 >  Other info:
 >
 >    $ grep -i time /var/run/dmesg.boot
 >    Timecounter "i8254"  frequency 1193182 Hz
 >    $ sysctl kern.timecounter
 >    kern.timecounter.method: 0
 >    kern.timecounter.hardware: i8254
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 It cannot possibly be much faster if such a slow timecounter is
 configured.  The i8254 timecounter uses 3 ISA bus accesses.  ISA bus
 accesses usually take even longer on n GHz machines now than they did
 on n MHz machines in 1992, since they are usually behind an ISA bridge
 or two and BIOSes have usually lost support for tweaking the number
 of ISA wait states to the mininum permitted by the spec and below.
 Some timings for reading the i8254 registers in a loop:
 
                            0x40 (i8254 mode)  0x43(i8254 counter 0)
 1996: (ASUS P55TP4XE)       703               1180 nsec
 1997: (FIC PA2007)         1105               1233
 2003: (ASUS A7V266/E)      1317               1317
 
 FreeBSD-4.8 should be configured to use the TSC timecounter in the non-SMP
 case.  Configuration of apm may pessimize the default configuration.  The
 TSC is not supported for the !SMP case.  Then the only alternative to the
 i8254 timecounter is the PIIX timecounter which AFAIK is only available on
 some old Intel chipsets (BX).  The PIIX timecounter hardware is a bit slower
 than the TSC.
 
 FreeBSD-5 can use the ACPI timecounter, which is essentially the same
 as the PIIX timecounter including its bugs.  FreeBSD-5 should be
 configured to use the TSC timecounter in the non-SMP case if it works.
 It doesn't work mainly if the CPU is ever throttled by power management.
 
 Some timings by another benchmark on an AthlonXP-1600 overclocked by
 146/133:
 
 %%%
 2002/01/15 4.5RC
 min 343, max 19573, mean 345.082019, std 109.709710
 1th: 344 (1354559 observations)
 2th: 343 (292357 observations)
 3th: 346 (275037 observations)
 4th: 345 (77690 observations)
 5th: 350 (33 observations)
 
 2002/01/15 5.0-2000/09/09
 min 343, max 40335, mean 346.613161, std 126.348898
 1th: 344 (1099112 observations)
 2th: 349 (635890 observations)
 3th: 343 (236910 observations)
 4th: 350 (27833 observations)
 5th: 352 (12 observations)
 
 2002/02/17 5.0-2002/02/17
 min 508, max 212269, mean 512.057268, std 96.120401
 1th: 509 (941667 observations)
 2th: 510 (608011 observations)
 3th: 517 (243539 observations)
 4th: 518 (123703 observations)
 5th: 508 (62574 observations)
 
 2002/02/17
 min 378, max 212621, mean 381.234928, std 115.229532
 1th: 380 (1139521 observations)
 2th: 381 (535937 observations)
 3th: 379 (263698 observations)
 4th: 378 (47663 observations)
 5th: 382 (12839 observations)
 
 2002/03/04
 min 378, max 237729, mean 379.740075, std 120.720743
 1th: 378 (1197389 observations)
 2th: 379 (802202 observations)
 3th: 380 (95 observations)
 4th: 476 (21 observations)
 5th: 397 (16 observations)
 
 2002/03/30 (ACPI-fast timecounter)
 min 837, max 210920, mean 892.293601, std 226.259588
 1th: 838 (1419633 observations)
 2th: 1117 (198530 observations)
 3th: 839 (183607 observations)
 4th: 1118 (169521 observations)
 5th: 837 (28327 observations)
 
 2002/03/30 (i8254 timecounter)
 min 4190, max 252275, mean 4301.709102, std 835.697703
 1th: 4191 (1113886 observations)
 2th: 4190 (683892 observations)
 3th: 5029 (148800 observations)
 4th: 5028 (51440 observations)
 5th: 28496 (662 observations)
 
 2002/03/30
 min 376, max 209046, mean 378.621313, std 116.589942
 1th: 378 (870256 observations)
 2th: 377 (470356 observations)
 3th: 376 (386644 observations)
 4th: 379 (263005 observations)
 5th: 380 (8563 observations)
 
 2003/06/04
 min 400, max 295020, mean 402.220277, std 101.969771
 1th: 401 (1078607 observations)
 2th: 400 (534769 observations)
 3th: 402 (329776 observations)
 4th: 403 (54325 observations)
 5th: 662 (317 observations)
 %%%
 
 Times are for the TSC timecounter except as noted.
 
 All times are in nsec.
 
 The benchmark uses clock_gettime() instead of the archaic, nonstandard,
 imprecise gettimeofday().  Old versions used getttimeofday() and broke
 a few years ago when average delta-time became less than 1 usec.
 
 Bruce