Intel DP55KG + Intel Core i7 870 + Phoronix Benchmarks / TurboBoost

From: Justin Piszcz
Date: Sun Nov 08 2009 - 17:47:49 EST

Next message: Russell King: "Re: [PATCH 22/23] sysctl arm: Remove binary sysctl support"
Previous message: Eric W. Biederman: "Re: [PATCH 22/23] sysctl arm: Remove binary sysctl support"
Next in thread: Justin Piszcz: "Re: Intel DP55KG + Intel Core i7 870 + Phoronix Benchmarks / TurboBoost"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hello,

I performed a bunch of tests with my processor to see if Turbo Boost was
working correct, the short answer is, it /appears/ to be..

CPU Tested: i870
Distribution: PTS Live CD 2009.3 x86_64
Distribution: Debian Testing x86_64

Intel's Benchmarks: (with a correctly working BIOS)
http://www.phoronix.com/scan.php?page=article&item=intel_lynnfield_add&num=4

Phoronix's Benchmarks: (when Turbo Boost is not working correctly)
http://www.phoronix.com/scan.php?page=article&item=intel_lynnfield&num=5

Note on PostgreSQL: This is disk-bound, so do not compare CD vs. NON-CD.

Phoronix Benchmark for Apache 2.2.11 (i870): 7422.11 static web page serving
Intel Benchmark for Apache 2.2.11 (i750): 8497.47 static web page serving
Same test my host (Phoronix Live CD) (i870): 10789.11 static web page serving*

Phoronix Benchmark for LAME 3.98.2 (i870): 46.01 seconds
Intel Benchmark for LAME 3.98.2 (i750): 26.18 seconds
Same test my host (Phoronix Live CD) (i870): 23.47 seconds*

Phoronix Benchmark for GraphicsMagick 1.3.6 (i870): 89.66 iterations
Intel Benchmark for GraphicsMagick 1.3.6 (i750): 124.66 iterations
Same test my host (Phoronix Live CD) (i870): 184.00 iterations*

Phoronix Benchmark for PostgreSQL 8.4.0 (i870): 540.87 transactions
Intel Benchmark for PostgreSQL 8.4.0 (i750): 1965.70 transactions
-
Same test my host (Phoronix Live CD) (i870): 7208.42 transactions*

--

w/no CPU_FREQ in kernel, running Phoronix test suite (pkg) on Debian/Testing
Kernel 2.6.31.4 x86_64 [keep in mind Phoronix Live CD runs on ramdisk], so
my main points of comparison will be [NO ACPI_FREQ vs. ACPI_FREQ]

Phoronix Benchmark for Apache 2.2.11 (i870): 7422.11 static web page
Intel Benchmark for Apache 2.2.11 (i750): 8497.47 static web page
Same test my host (Phoronix Live CD) (i870): 10789.11 static web page
Same test my host (Debian w/NO ACPI_FREQ) (i870): 12026.90 static web page*
Same test my host (Debian w/ACPI_FREQ) (i870): 11883.58 static web page-

Phoronix Benchmark for LAME 3.98.2 (i870): 46.01 seconds
Intel Benchmark for LAME 3.98.2 (i750): 26.18 seconds
Same test my host (Phoronix Live CD) (i870): 23.47 seconds*
Same test my host (Debian w/NO ACPI_FREQ) (i870): 29.82 seconds-
Same test my host (Debian w/ACPI_FREQ) (i870): 26.36 seconds+

Phoronix Benchmark for GraphicsMagick 1.3.6 (i870): 89.66 iterations
Intel Benchmark for GraphicsMagick 1.3.6 (i750): 124.66 iterations
Same test my host (Phoronix Live CD) (i870): 184.00 iterations*
Same test my host (Debian w/NO ACPI_FREQ) (i870): 152.33 iterations-
Same test my host (Debian w/ACPI_FREQ) (i870): 165.66 iterations+

Phoronix Benchmark for PostgreSQL 8.4.0 (i870): 540.87 transactions
Intel Benchmark for PostgreSQL 8.4.0 (i750): 1965.70 transactions
Same test my host (Phoronix Live CD) (i870): 7208.42 transactions*
-
Same test my host (Debian w/NO ACPI_FREQ) (i870): 1446.41 transactions+
Same test my host (Debian w/ACPI_FREQ) (i870): 1417.55 transactions-

--

From my bzip2 test:

Phoronix Live CD:
ptslive@Gernlinden:~$ ./run_test.sh
39.81user 0.10system 0:39.91elapsed 100%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+2025minor)pagefaults 0swaps

My host:
50.14user 0.11system 0:50.27elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+2653minor)pagefaults 0swaps

10 seconds faster! Question is-- is Ubuntu's distribution that much faster?
Or is it just the way the kernel is configured?

--

ok, so is it the Live CD / Distribution itself that lends itself to the speed
or is it the kernel configuration? took the config-2.6.31-3 from the Live CD's
/boot and used that as my kernel base as a test:

--

First, a quick test: (bzip2 -9 2.6.31 kernel tarball)

bzip2 test Debian vs. Ubuntu, same kernel configuration (minus 2.6.31 vs.
2.6.31.4)

My host (w/phoronix kernel configuration)

Almost 10 seconds faster:
40.97user 0.09system 0:41.07elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
64inputs+0outputs (1major+2021minor)pagefaults 0swaps

.. interesting .. hopefully the rest of the benchmarks (4) will reflect faster
.. speeds as well ..

--

My kernel .config vs. Phoronix's Distribution's .config: (yes I realize
glibc/etc/libraries could be all different versions, etc-- but I want to see
if this makes a difference), e.g. the Phoronix kernel uses NO_HZ, 250hz, much
different than my configuration... And also, Phoronix's kernel is 2.6.31, th
one I am using is 2.6.31.4.

Intel Benchmark for Apache 2.2.11 (i750): 8497.47 static web page
Same test my host (Phoronix Live CD) (i870): 10789.11 static web page
Same test my host (Debian w/ACPI_FREQ) (i870): 11883.58 static web page
Same test my host (Debian w/PHORONIX_CFG) (i870): 9343.15 static web page

- slower here, is this due to 250hz vs 1000hz setting (?)

Intel Benchmark for LAME 3.98.2 (i750): 26.18 seconds
Same test my host (Phoronix Live CD) (i870): 23.47 seconds*
Same test my host (Debian w/ACPI_FREQ) (i870): 26.36 seconds+
Same test my host (Debian w/PHORONIX_CFG) (i870): 24.24 seconds-

Intel Benchmark for GraphicsMagick 1.3.6 (i750): 124.66 iterations
Same test my host (Phoronix Live CD) (i870): 184.00 iterations*
Same test my host (Debian w/ACPI_FREQ) (i870): 165.66 iterations-
Same test my host (Debian w/PHORONIX_CFG) (i870): 173.66 iterations+

Intel Benchmark for PostgreSQL 8.4.0 (i750): 1965.70 transactions
Same test my host (Phoronix Live CD) (i870): 7208.42 transactions*
-
Same test my host (Debian w/ACPI_FREQ) (i870): 1417.55 transactions
Same test my host (Debian w/PHORONIX_CFG) (i870): 1466.55 transactions

--

Phoronix kernel config from live CD:
http://home.comcast.net/~jpiszcz/20091108/config-2.6.31-3-generic.phoronix.txt

My kernel config:
http://home.comcast.net/~jpiszcz/20091108/config-2.6.31.4.txt

Conclusion: Using the +ACPI_FREQ option yields some advantage; however,
it does make the CPU/motherboard emit a very high pitch noise when the
system is not doing anything. The phoronix kernel config with +ACPI_FREQ enabled and the kernel options they set yield better performance than my kernel configuration. The next question is which options (outside of ACPI_FREQ) yield the +10 seconds faster speed with bzip2 (along with the other 3 benchmarks outside of Apache, which was slower).

Just modified my kernel configuration:

My regular kernel configuration (does not have NO_HZ or ACPI_FREQ)
50.14user 0.11system 0:50.27elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+2653minor)pagefaults 0swaps

ACPI_FREQ+NO_HZ+100 HZ
40.93user 0.15system 0:41.17elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1major+2020minor)pagefaults 0swaps

ACPI_FREQ+NO_HZ+250 HZ
41.04user 0.14system 0:41.24elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1major+2021minor)pagefaults 0swaps

ACPI_FREQ+NO_HZ+300 HZ
40.89user 0.11system 0:41.11elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1major+2023minor)pagefaults 0swaps

ACPI_FREQ+NO_HZ+1000 HZ
41.62user 0.10system 0:41.80elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1major+2021minor)pagefaults 0swaps

ACPI_FREQ+1000 HZ (but disable NO_HZ) - almost back to ~50 seconds again
44.58user 0.15system 0:44.77elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1major+2022minor)pagefaults 0swaps

NO_HZ+1000 HZ (but disable ACPI_FREQ) - back to ~50 seconds
50.00user 0.10system 0:50.16elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1major+2020minor)pagefaults 0swaps

So NO_HZ accompanied by ACPI_FREQ speeds up the bzip2 by ~10 seconds (below):
41.20user 0.11system 0:41.37elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1major+2022minor)pagefaults 0swaps

More tests are needed but I was solely interested in single-core performance to take advantage of turbo boost. Comparing the results on the phoronix live CD with those from Intel when they benchmarked the i750, they seem to be in-line with what one would expect (overall)..

Justin.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Next message: Russell King: "Re: [PATCH 22/23] sysctl arm: Remove binary sysctl support"
Previous message: Eric W. Biederman: "Re: [PATCH 22/23] sysctl arm: Remove binary sysctl support"
Next in thread: Justin Piszcz: "Re: Intel DP55KG + Intel Core i7 870 + Phoronix Benchmarks / TurboBoost"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]