3. One Page Summary
Expensive middleware:
• is priced per core differently for different processors;
• its costs overshadow server & operating system costs.
1. Assess using Performance per Weighted Core
(P/WC) rather than Performance per Core.
2. The observed best to worst ratio exceeds 7x in one case.
3. Results vary not only with middleware and application,
but also with the scale.
4. For TPC-C on Oracle DBMS, the Hyper-threaded
XEONs have the best:
• P/WC rating;
• Clock-independent architecture:
P/WC per GHz.
5. For TPC-C on Oracle, similar results come from using
IBM’s PVU rather than Oracle’s Core Weighting Factor.
See below for
TPC-H and SAP
results as well.
3
4. Processor Selection For Optimum Middleware Price / Performance
One Page Summary
dakra137 at gmail.com
Expensive middleware:
• is priced per core differently for
different processors, servers,
etc;
• its costs overshadow server &
operating system costs.
1. Assess using Performance per
Weighted Core (P/WC) rather
than Performance per Core.
2. The observed best to worst ratio
exceeds 7x in one case.
3. Results vary not only with
middleware and application, but
also with the scale.
4. For TPC-C on Oracle DBMS, the
Hyper-threaded XEONs have the
best:
• P/WC rating;
• Clock rate-independent
architecture:
P/WC per GHz.
5. For TPC-C on Oracle, similar
results come from using IBM’s
PVU rather than Oracle’s Core
Weighting Factor.
6. For TPC-H on Oracle:
• Itanium
• Top @ 10,000
• Bottom @ 3,000 and 1,000
• Sparc
• Top (along with Power) @
3,000
• 2nd @ 10,000 and 1,000
• Power
• Top @ 3,000
• Hyper-threaded Xeon
• Top @ 1,000
7. SAP on Oracle
• Hyper-threaded Xeon on top
8. SAP on DB2
• Power7 on top
4
5. Acknowledgements
I wish to thank:
• The Transaction Processing Council and
SAP for making it so easy to acquire
benchmark results in readily usable form;
• Oracle and IBM for making their Core
Weighting Factor and Processor Value
Unit Tables readily available;
and especially,
• All the staff at all the companies who
implemented, tuned, measured, and
published the TPC and SAP benchmarks. 5
6. Key Question
How do you take into account
“Core Weighting Factors,”
“Processor Value Units,” etc.
when looking to pick a
processor to run EXPEN$IVE
middleware?
6
11. Answer #2a
That really does depend on:
• the middleware,
• the application, and
• its scale.
11
12. Answer #2b
This case study
considers:
• Middleware:
Oracle DBMS
• Applications:
•TPC-C
•TPC-H
TPC-H scales:
•1,000
•3,000
•10,000 12
13. Key Points
Middleware $ >> Everything else that
varies e.g. Server
HW & OS
This case study avoided:
• Exact pricing, for the reason above
• Performing benchmarks
• Constraints on publishing results
13
17. Key Point
Don’t be mislead by the quantities of
benchmarks!
The quantity of entries represents the
willingness of vendors to perform
benchmarks and publish results. All it
shows is consistency even as other
elements may vary, such as chipset,
SAN attachment, storage subsystem, etc.17
19. Architecture Analysis
For TPC-C the Hyper-threaded XEONs come out on top
because of performance, advantageous core weighting
factor, and clock rate.
The Power 6’s are next, due to sheer clock rate, in spite of heaviest
weighting.
The SPARC T3 would still be in the top 20 chart, ahead of the
some of the XEONs, even if it had the same weighting factor (.5
rather than .25) as the XEONs.
The best Itanium result compares closely to the Power5 results.
While the best XEON rating is 158K, the Itanium results vary from
58K down to 45K TpMC/WC rating units.
19
20. Source of Top Xeon’s Rating
Cache? No, the top three outperformed the
Xeon’s with larger cache per core and larger
total cache. 2MB per core was enough here.
Clock rate? No, the top three
outperformed the Xeon’s with higher clock
rates.
Hyper-threading? YES!
Core Weighting Factor? YES!
20
25. Analysis of TPC-H Results
TPC-H Scale Range of Results
10,000 1:2.4
3,000 1:5.6
1,000 1:7.4
Selecting a suboptimal processor
for TPC-H could increase costs
by up to 7.4x.
25
26. Analysis of TPC-H Results
Scale really matters!
• Itanium
• Top @ 10,000
• Bottom @ 3,000 and 1,000
• Sparc
• Top (along with Power) @ 3,000
• 2nd @ 10,000 and 1,000
• Greater gaps between groups of results @ 10,000 and 1,000
than @ 3,000
26
32. Clock Rate Will Grow
Compare architectures.
Factor out clock rate, so you can
estimate new higher clock rate
chips’ performance.
Use:
Performance per Weighted
Core per GHz 32
37. TPC-C Ranking Shifts Due to
Alternative Weightings
Server CPU Type
Order due to Oracle
Core Weighting
Factors
Order due to IBM
Processor Value
Units
Order
Difference
Intel Xeon Processor X5570 2.93GHz 1 1 0
Intel Xeon E5520 2.27 GHz 3 4 -1
IBM POWER6 - 4.7 GHz 4 3 1
Intel Xeon X5650 6-core 2.66GHz 5 5 0
IBM POWER5+ - 2.2 GHz 6 7 -1
SPARC T3 1.65GHz 7 31 -24
Intel Xeon Quad-Core X5460 - 3.16 GHz 8 6 2
IBM POWER5 - 1.9 GHz 9 12 -3
IBM POWER5 - 1.9 GHz 10 15 -5
IBM POWER5 - 1.9 GHz 12 18 -6
Intel Itanium2 Dual-Core - 1.6 GHz 13 8 5
Intel Xeon X7460 - 2.67 GHz 14 9 5
Intel Xeon X5355 - 2.66 GHz 17 13 4
Intel Itanium2 Dual-Core - 1.6 GHz 18 14 4
Intel Xeon QC 5440 - 2.83 GHz 19 17 2
Similar rows omitted 37
38. Different Architecture
Q: What will core weighting
factors and PVU’s be for the new
AMD Bulldozer processors?
(They are somewhat less than 2 cores per
processor building block.)
A: I don’t know.
38