xref: /freebsd/share/doc/papers/diskperf/results.ms (revision 1e413cf93298b5b97441a21d9a50fdcd0ee9945e)
Copyright (c) 1983 The Regents of the University of California.
All rights reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
3. All advertising materials mentioning features or use of this software
must display the following acknowledgement:
This product includes software developed by the University of
California, Berkeley and its contributors.
4. Neither the name of the University nor the names of its contributors
may be used to endorse or promote products derived from this software
without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
SUCH DAMAGE.

@(#)results.ms 6.2 (Berkeley) 4/16/91

Results

The following tables indicate the results of our test runs. Note that each table contains results for tests run on two varieties of 4.2BSD file systems. The first set of results is always for a file system with a basic blocking factor of eight Kilobytes and a fragment size of 1 Kilobyte. The second sets of measurements are for file systems with a four Kilobyte block size and a one Kilobyte fragment size. The values in parenthesis indicate the percentage of CPU time used by the test program. In the case of the two disk arm tests, the value in parenthesis indicates the sum of the percentage of the test programs that were run. Entries of ``n. m.'' indicate this value was not measured.

4.2BSD File Systems Tests - VAX 11/750
Logically Sequential Transfers
from an 8K/1K 4.2BSD File System (Kbytes/sec.)
Test Emulex SC750/Eagle UDA50/RA81
1 Drive 2 Drives 1 Drive 2 Drives
read_8192 490 (69%) 620 (96%) 310 (44%) 520 (65%)
write_4096 380 (99%) 370 (99%) 370 (97%) 360 (98%)
write_8192 470 (99%) 470 (99%) 320 (71%) 410 (83%)
rewrite_8192 650 (99%) 620 (99%) 310 (50%) 450 (70%)
.T&
c s s s s
c s s s s
l | l s | l s
l | l s | l s
l | l l | l l
l | c c | c c.
Logically Sequential Transfers
from 4K/1K 4.2BSD File System (Kbytes/sec.)
Test Emulex SC750/Eagle UDA50/RA81
1 Drive 2 Drives 1 Drive 2 Drives
read_8192 300 (60%) 400 (84%) 210 (42%) 340 (77%)
write_4096 320 (98%) 320 (98%) 220 (67%) 290 (99%)
write_8192 340 (98%) 340 (99%) 220 (65%) 310 (98%)
rewrite_8192 450 (99%) 450 (98%) 230 (47%) 340 (78%)

Note that the rate of write operations on the VAX 11/750 are ultimately CPU limited in some cases. The write rates saturate the CPU at a lower bandwidth than the reads because they must do disk allocation in addition to moving the data from the user program to the disk. The UDA50/RA81 saturates the CPU at a lower transfer rate for a given operation than the SC750/Eagle because it causes more memory contention with the CPU. We do not know if this contention is caused by the UNIBUS controller or the UDA50.

The following table reports the results of test runs on a VAX 11/780 with 4 Megabytes of main memory.

4.2BSD File Systems Tests - VAX 11/780
Logically Sequential Transfers
from an 8K/1K 4.2BSD File System (Kbytes/sec.)
Test Emulex SC780/Eagle UDA50/RA81 Sys. Ind. 9900/Eagle
1 Drive 2 Drives 1 Drive 2 Drives 1 Drive 2 Drives
read_8192 560 (70%) 480 (58%) 360 (45%) 540 (72%) 340 (41%) 520 (66%)
write_4096 440 (98%) 440 (98%) 380 (99%) 480 (96%) 490 (96%) 440 (84%)
write_8192 490 (98%) 490 (98%) 220 (58%)* 480 (92%) 490 (80%) 430 (72%)
rewrite_8192 760 (100%) 560 (72%) 220 (50%)* 180 (52%)* 490 (60%) 520 (62%)
.T&
c s s s s s s
c s s s s s s
l | l s | l s | l s
l | l s | l s | l s
l | l l | l l | l l
l | c c | c c | c c.
Logically Sequential Transfers
from an 4K/1K 4.2BSD File System (Kbytes/sec.)
Test Emulex SC780/Eagle UDA50/RA81 Sys. Ind. 9900/Eagle
1 Drive 2 Drives 1 Drive 2 Drives 1 Drive 2 Drives
read_8192 490 (77%) 370 (66%) n.m. n.m. 200 (31%) 370 (56%)
write_4096 380 (98%) 370 (98%) n.m. n.m. 200 (46%) 370 (88%)
write_8192 380 (99%) 370 (97%) n.m. n.m. 200 (45%) 320 (76%)
rewrite_8192 490 (87%) 350 (66%) n.m. n.m. 200 (31%) 300 (46%)
* the operation of the hardware was suspect during these tests.

The dropoff in reading and writing rates for the two drive SC780/Eagle tests are probably due to the file system using insufficient rotational delay for these tests. We have not fully investigated these times.

The following table compares data rates on VAX 11/750s directly with those of VAX 11/780s using the UDA50/RA81 storage system.

4.2BSD File Systems Tests - DEC UDA50 - 750 vs. 780
Logically Sequential Transfers
from an 8K/1K 4.2BSD File System (Kbytes/sec.)
Test VAX 11/750 UNIBUS VAX 11/780 UNIBUS
1 Drive 2 Drives 1 Drive 2 Drives
read_8192 310 (44%) 520 (84%) 360 (45%) 540 (72%)
write_4096 370 (97%) 360 (100%) 380 (99%) 480 (96%)
write_8192 320 (71%) 410 (96%) 220 (58%)* 480 (92%)
rewrite_8192 310 (50%) 450 (80%) 220 (50%)* 180 (52%)*
.T&
c s s s s
c s s s s
l | l s | l s
l | l s | l s
l | l l | l l
l | c c | c c.
Logically Sequential Transfers
from an 4K/1K 4.2BSD File System (Kbytes/sec.)
Test VAX 11/750 UNIBUS VAX 11/780 UNIBUS
1 Drive 2 Drives 1 Drive 2 Drives
read_8192 210 (42%) 342 (77%) n.m. n.m.
write_4096 215 (67%) 294 (99%) n.m. n.m.
write_8192 215 (65%) 305 (98%) n.m. n.m.
rewrite_8192 227 (47%) 336 (78%) n.m. n.m.
* the operation of the hardware was suspect during these tests.

The higher throughput available on VAX 11/780s is due to a number of factors. The larger main memory size allows a larger file system cache. The block allocation routines run faster, raising the upper limit on the data rates in writing new files.

The next table makes the same comparison using an Emulex controller on both systems.

4.2BSD File Systems Tests - Emulex - 750 vs. 780
Logically Sequential Transfers
from an 8K/1K 4.2BSD File System (Kbytes/sec.)
Test VAX 11/750 CMI Bus VAX 11/780 SBI Bus
1 Drive 2 Drives 1 Drive 2 Drives
read_8192 490 (69%) 620 (96%) 560 (70%) 480 (58%)
write_4096 380 (99%) 370 (99%) 440 (98%) 440 (98%)
write_8192 470 (99%) 470 (99%) 490 (98%) 490 (98%)
rewrite_8192 650 (99%) 620 (99%) 760 (100%) 560 (72%)
.T&
c s s s s
c s s s s
l | l s | l s
l | l s | l s
l | l l | l l
l | c c | c c.
Logically Sequential Transfers
from an 4K/1K 4.2BSD File System (Kbytes/sec.)
Test VAX 11/750 CMI Bus VAX 11/780 SBI Bus
1 Drive 2 Drives 1 Drive 2 Drives
read_8192 300 (60%) 400 (84%) 490 (77%) 370 (66%)
write_4096 320 (98%) 320 (98%) 380 (98%) 370 (98%)
write_8192 340 (98%) 340 (99%) 380 (99%) 370 (97%)
rewrite_8192 450 (99%) 450 (98%) 490 (87%) 350 (66%)

The following table illustrates the evolution of our testing process as both hardware and software problems effecting the performance of the Emulex SC780 were corrected. The software change was suggested to us by George Goble of Purdue University.

The 4.2BSD handler for RH750/RH780 interfaced disk drives contains several constants which to determine how much time is provided between an interrupt signaling the completion of a positioning command and the subsequent start of a data transfer operation. These lead times are expressed as sectors of rotational delay. If they are too small, an extra complete rotation will often be required between a seek and subsequent read or write operation. The higher bit rate and rotational speed of the 2351A Fujitsu disk drives required increasing these constants.

The hardware change involved allowing for slightly longer delays in arbitrating for cycles on the SBI bus by starting the bus arbitration cycle a little further ahead of when the data was ready for transfer. Finally we had to increase the rotational delay between consecutive blocks in the file because the higher bandwidth from the disk generated more memory contention, which slowed down the processor.

4.2BSD File Systems Tests - Emulex SC780 Disk Controller Evolution
Logically Sequential Transfers
from an 8K/1K 4.2BSD File System (Kbytes/sec.)
Test Inadequate Search Lead OK Search Lead OK Search Lead
Initial SBI Arbitration Init SBI Arb. Improved SBI Arb.
1 Drive 2 Drives 1 Drive 2 Drives 1 Drive 2 Drives
read_8192 320 370 440 (60%) n.m. 560 (70%) 480 (58%)
write_4096 250 270 300 (63%) n.m. 440 (98%) 440 (98%)
write_8192 250 280 340 (60%) n.m. 490 (98%) 490 (98%)
rewrite_8192 250 290 380 (48%) n.m. 760 (100%) 560 (72%)
.T&
c s s s s s s
c s s s s s s
l | l s | l s | l s
l | l s | l s | l s
l | l s | l s | l s
l | c c | c c | c c
l | c c | c c | c c.
Logically Sequential Transfers
from an 4K/1K 4.2BSD File System (Kbytes/sec.)
Test Inadequate Search Lead OK Search Lead OK Search Lead
Initial SBI Arbitration Init SBI Arb. Improved SBI Arb.
1 Drive 2 Drives 1 Drive 2 Drives 1 Drive 2 Drives
read_8192 200 220 280 n.m. 490 (77%) 370 (66%)
write_4096 180 190 300 n.m. 380 (98%) 370 (98%)
write_8192 180 200 320 n.m. 380 (99%) 370 (97%)
rewrite_8192 190 200 340 n.m. 490 (87%) 350 (66%)
.bp