• Choose a category
• All
• Classic-Fiction

* * *

[3] On FreeBSD, softupdates aren't a mount option, but must be set with the tunefs command.

OpenBSD

The results in this section are from OpenBSD Version 3.3 (released May 1, 2003). I built a kernel with the following notable configuration options:

option MSGMNB=8192

option MSGMNI=40

option MSGSEG=512

option MSGSSZ=64

option MSGTQL=2048

option SHMSEG=16

option SHMMNI=32

option SHMMAX=2048

option SHMALL=4096

option NMBCLUSTERS=32768

option MAXFILES=8192

Table D-3 and Figure D-3 summarize the OpenBSD results. The choices for OpenBSD are similar to those for FreeBSD. Unfortunately, however, coss doesn't run on OpenBSD, which lacks the aio_read( ) and aio_write( ) functions.

Table D-3. OpenBSD benchmarking results

Storage scheme

Filesystem

Mount options

Throughput

Response time

Hit ratio

diskd(1)

UFS

async, noatime, softupdate

91.1

1.45

56.3

diskd(2)

UFS

63.7

1.44

56.2

ufs(1)

UFS

softupdate

27.6

1.51

56.3

ufs(2)

UFS

noatime

25.1

1.52

56.3

ufs(3)

UFS

22.7

1.52

56.1

ufs(4)

UFS

async

22.1

1.51

56.6

Figure D-3. OpenBSD filesystem benchmarking traces

In general, the OpenBSD results are slightly worse than FreeBSD. This isn't too surprising, given that the OpenBSD project emphasizes security and perhaps spends less time on filesystem performance.

One odd result is that using the async option (alone) caused a slight decrease in performance for the ufs storage scheme.

NetBSD

These results come from NetBSD Version 1.6.1 (released April 21, 2003). Table D-4 and Figure D-4 summarize the NetBSD results. NetBSD actually performs almost the same as OpenBSD. The best configuration yields about 90 transactions per second. Unfortunately, NetBSD doesn't support coss or aufs. I built a custom kernel with these options:

options NMBCLUSTERS=32768

options MAXFILES=8192

options MSGSSZ=64

options MSGSEG=512

options MSGMNB=8192

options MSGMNI=40

options MSGTQL=2048

Table D-4. NetBSD benchmarking results

Storage scheme

Filesystem

Mount options

Throughput

Response time

Hit ratio

diskd(1)

UFS

softupdate,noatime,async

90.3

1.49

57.2

diskd(2)

UFS

softupdate

73.5

1.51

55.8

diskd(3)

UFS

60.1

1.48

55.9

ufs(1)

UFS

softupdate,noatime,async

34.9

1.51

56.2

ufs(2)

UFS

softupdate

31.7

1.52

55.5

ufs(3)

UFS

23.6

1.53

55.4

Figure D-4. NetBSD filesystem benchmarking traces

Solaris

These results come from Solaris Version 8 for Intel (released February 2002). Solaris 9 was available when I started these tests, but Sun no longer makes it freely available. I tweaked the kernel by adding these lines to /etc/system:

set rlim_fd_max = 8192

set msgsys:msginfo_msgmax=8192

set msgsys:msginfo_msgmnb=8192

set msgsys:msginfo_msgmni=40

set msgsys:msginfo_msgssz=64

set msgsys:msginfo_msgtql=2048

set shmsys:shminfo_shmmax=2097152

set shmsys:shminfo_shmmni=32

set shmsys:shminfo_shmseg=16

Table D-5 and Figure D-5 summarize the Solaris results. This is the only other operating system, in addition to Linux, in which the aufs storage scheme works well. Interestingly, both aufs and diskd have about the same performance on Solaris, although the actual numbers are much lower than on Linux.

Table D-5. Solaris benchmarking results

Storage scheme

Filesystem

Mount options

Throughput

Response time

Hit ratio

diskd(1)

UFS

noatime

56.3

1.53

55.7

aufs(1)

UFS

noatime

53.6

1.49

56.6

diskd(2)

UFS

37.9

1.53

55.5

aufs(2)

UFS

37.4

1.49

56.4

coss

32.4

1.47

54.6

ufs(1)

UFS

noatime

24.0

1.53

55.6

ufs(2)

UFS

19.0

1.50

56.3

Figure D-5. Solaris filesystem benchmarking traces

Solaris also supports coss, but at nowhere near the rates for Linux and FreeBSD. For some unknown reason, coss on Solaris is limited to 32 transactions per second.

Number of Disk Spindles

In this section, I compare Squid's performance for different number of disk drives (spindles). These tests are from the Linux system with the aufs storage scheme and ext2fs filesystems.