(www.opensound.com/oss.html) because of a lack of
reliable support for full duplex with the test system
hardware. Table 5 shows the results of the tests with system
load.
System Operating Details Latency
System (ms)
Spirit - - 1.81
A Linux 2.4.1 (AM) ALSA 2.72
A Linux 2.4.1 ALSA 2.72
F Windows 2000 ASIO 3.11
C MacOS X CoreAudio 3.97
F Linux 2.4.1 (IM) ALSA (L) 4.30
F Linux 2.4.5 (AM) ALSA (L) 4.30
F Linux 2.4.2 ALSA (L) 4.30
G MacOS 9.04 ASIO 6.80
A Linux 2.2.16 ALSA 7.19
F Windows 2000 MME (P) 11.45
E Linux 2.4.0 OSS 12.20
E Windows 98 MME (P) 60.86
E Windows 98 DirectSound (P) 63.24
D Windows ME MME (P) 73.51
A Windows 2000 MME (P) 75.85
D Windows ME DirectSound (P) 82.86
B MacOS 8.6 SM (P) 106.24
A Windows 2000 DirectSound (P) 123.11
C MacOS 9.04 SM (P) VM Off 195.58
C MacOS 9.04 SM (P) VM On 935.53
Table 4. Latency Test results for systems without load.
P - PortAudio. L - LAAGA. VM - MacOS virtual memory
settings. AM - Andrew Morton. IM - Ingo Molnar.
System Operating System Details Latency (ms)
C MacOS X CoreAudio 3.97
F Linux 2.4.1 (IM) ALSA (L) 4.30
F Linux 2.4.5 (AM) ALSA (L) 4.30
F Linux 2.4.2 ALSA (L) 4.30
F Windows 2000 ASIO 6.03
G MacOS 9.04 ASIO 6.80
F Windows 2000 MME (P) 245.17
Table 5 - Latency Test Results with system load.
P - PortAudio. L - LAAGA.
AM - Andrew Morton. IM - Ingo Molnar.
5.1 Linux
The Linux systems had the best performance on the tests
without load and the second best performance with load.
Furthermore, the Linux systems tended to have the least
difference between best-case and worst-case performance.
Excluding the test that used the less efficient OSS API,
which is slated for replacement by ALSA (Advanced Linux
Sound Architecture) (www.alsa-project.org) in the next
kernel version, all of the Linux tests had latencies below 10
milliseconds. Finally, the performance of the Linux systems
are independent of the quality and expense of the
soundcards.
The tests using the kernels with low-latency
enhancements did not show significantly better performance
than the standard kernels. This is most likely due to the test
hardware used and the simplicity of the load generation
programs. The low-latency patches address issues only with
specific kernel sub-systems rather than providing a general
enhancement to the latency characteristics of the system. It
is likely that the load generation programs used did not
exercise the improved sub-systems. Additionally, in our
experience, most of the latency problems in Linux are
related to IDE disks, which none of the systems tested with
load used. Performing other tasks with the systems while
running the latency test program, such as CPU intensive
graphics programs and compiling large projects, showed
that some disk-intensive tasks can cause problems with the
standard kernel. No amount of CPU load alone created
latency problems, however.
One interesting aspect of these results is the excellent
latency performance using LAAGA. The only other results
that were less than 10 milliseconds were ASIO (Audio
Streaming Input Output) (Steinberg 1999) on Windows and
Macintosh, neither of which allow more than one
application access to the sound hardware concurrently as
LAAGA allows. LAAGA showed excellent performance
that did not noticeably degrade with multiple test
applications accessing the sound hardware concurrently.
5.2 Microsoft Windows
The Windows systems tested had results ranging from
excellent to poor in correlation with the quality of the
soundcard being tested. All of the tests using the RME
Hammerfall card without system load performed well, but
all of the tests with consumer-quality soundcards had
latencies above 60 milliseconds. For the tests with system
load, only the ASIO drivers performed well. The test
program using the MME (Multimedia Extensions) (Petzold
1998) driver performed poorly with load. Preliminary tests
showed that this was caused more by CPU load than disk
load.
5.3 Apple MacOS 8 and 9
Like the Windows systems, the classic MacOS systems
had both excellent and poor performance, again correlated
with soundcard quality. The performance with ASIO, which
was tested using Max/MSP, was excellent in both the loaded
and unloaded tests. One surprising result was the latency
test using MacOS 9.04 with virtual memory enabled, which
yielded latencies of nearly 1 second.
5.4 Apple MacOS X
Apple MacOS X was the second best performing system
on the tests without load and the best performing system on
the tests with load. The new CoreAudio API, which was
introduced with MacOS X (Apple 2001), is clearly designed
0
