FORM-2
THE PATENTS ACT, 1970
(39 of 1970)
&
THE PATENTS RULES, 2003
PROVISIONAL
Specification
(See section 10 and rule 13)
ESTIMATION OF PERFORMANCE PARAMETERS FOR CENTRAL PROCESSING UNITS (CPUs) IN A SYSTEM
TATA CONSULTANCY SERVICES LTD.
an Indian Company
of Bombay House, 24, Sir Homi Mody Street, Mumbai 400 001,
Maharashtra, India,
THE FOLLOWING SPEC IFICATION DESCRIBES THE INVENTION.

Field of Invention:
The invention relates to the estimation of performance parameters for Central Processing Units (CPUs) in a system.
Background of Invention:
TPC-C is the industry standard for OLTP (online transaction processing) performance ratings of servers. Server vendors publish their system ratings in the form of tpmC (transactions per minute under TPC-C). Similarly, SPECjAppServer2004 is the industry standard for J2EE Application Server performance. The vendors publish their system ratings in the form of JOPS (J2EE business operations per second under SPECjAppServer2004 specifications).
TPC and JOPS are useful metrics for capacity planning of a system. There are several challenges in determining the TPC or JOPS rating of a CPU like:
• Vendors are very liberal in system and storage configuration during estimating TPC-C and SPEC benchmarks and the system is also finely tuned to maximize the result of the benchmark.
• To determine the TPC rating of a CPU which has no published TPC-C results, e.g. SUN Ultra Sparc IV+ or SUN Niagra. To determine the JOPS rating of a CPUs which do not have published SPEC results like IBM Power 4.
• To interpolate a TPC or JOPS rating of a CPU that is different from published results either in architecture or in Ghz. For example, an IBM Power 4 tpmC rating when all that is available are Power 5 tpmC
2

ratings, or an IBM Power 5 at 1.6 Ghz that you have and available is ratings for IBM Power 5 at 1.9 Ghz.
• To extrapolate ratings for a different operating system. For example, to determine the TPC and JOPS rating while running Red Hat Linux from the ratings for running on AMD Opteron under Solaris.
Thus there was a need for a TPCMeter and JOPSMeter to estimate the tpmC and JOPS rating of a CPU by means of running a simple test on the system at hand.
Objects of invention:
The object of this invention is to provide a tool for estimating the TPC and JOPS ratings of a CPU.
Another object of this invention is to provide a tool which will determine the TPC and JOPS rating for different operating systems.
Still one more object of this invention is to provide a tool to estimate the TPC and JOPS ratings of a CPU by running a simple test on the system, that completes in less than a minute with no sophisticated software setup other than the operating system, and without the need for enterprise class servers to host the CPUs.
3

Summary of the Invention:
The tool in accordance with this invention envisages a means for measuring the TPC and JOPS ratings of a CPU.
In accordance with one practical embodiment of this invention, the TPC meter and JOPS meter run on a system to and estimate the tpmC and the JOPS rating of the CPU, regardless of the storage, memory, and network. The tools are written in C and need a standard compiler, with a requirement of 1GB disk space and at least 512MB of RAM.
In accordance with one practical embodiment of this invention, to eliminate the effect of the network on the estimation of the ratings, the TPC meter tool runs on the server that is being calibrated. The RAM used while calibrating is significantly lower so as to eliminate the effect of RAM. To eliminate the effect of storage, no physical I/Os are carried out.
In accordance with one practical embodiment of this invention, a TPC transaction, on an average carries out 16 reads, 14.3 writes and 14.3/8 sequential writes, in terms of page accesses. Thus there are 49.9% random reads, 45.5% random writes and 5.6% sequential writes.
In accordance with one embodiment of this invention, the file size is kept very small, typically 128MB so that there is no physical IO when operations are in progress.
4

In accordance with one embodiment of this invention, for random writes the implementation is done by seeking a random position in the file and writing only one block at a time and average rate for it is determined. Using the weights obtained, rates for random reads, random writes, and sequential writes, are averaged to yield a net number for IOs per second. Then a CPU whose tpmC rating is known is taken and is used as the reference tpmC. The IOs per second obtained through the program is then normalized by dividing with a constant to match the reference tpmC.
In accordance with one practical embodiment of this invention, the JOPS and tpmC ratings are proportional. The JOPS meter essentially takes the IO rate calculated by TPC meter and divides by a normalizing factor to match a reference JOPS. The normalization factor then becomes a constant to estimate JOPS of all other CPUs.
Detailed Description of the Invention:
According to this invention there is provided a tool for measuring the TPC and JOPS ratings of a CPU.
The TPC meter and JOPS meter can run on any system and estimate the tpmC and the JOPS rating of the CPU, regardless of the storage, memory, and network. The tools are written in C and can be complied using any standard compiler with a requirement of 1GB disk space and at least 512MB of RAM.
5

The environment for estimating the tpmC and JOPS ratings using the present invention is not same as the environment a vendor of the CPU has for TPC-C benchmark. The CPU in both the cases is same, but there is a difference in system storage, RAM, and network. To eliminate the effect of the network, the TPC meter tool is run on the server that is being calibrated. The RAM used while calibrating is significantly lower so as to eliminate the effect of RAM. To eliminate the effect of storage, no physical I/Os are carried out.
A TPC transaction, on an average carries out 16 read, 6.3 writes and 8 updates. Treating a write and update as equal there are 16 reads and 14.3 writes. These are random IOs. For every write there is an equivalent sequential to the log file. The write to the log file if considered in blocks or IOs would essentially club records that fit into a page. The page size is 8KB and has 8 records per page for sequential IO. Thus total sequential writes are 14.3/8. Overall there are 16 reads, 14.3 writes and 14.3/8 sequential writes. Thus there are 49.9% random reads, 45.5% random writes and 5.6% sequential writes.
For the purpose of estimation of the rate at which the CPU can perform random and sequential reads and writes, the file size is kept very small, typically 128MB so that there is no physical IO when operations are in progress. The files are written sequentially and then it is repeated for a small number of writes and the test is repeated to get average values.
For random writes the implementation is done by seeking a random position in the file and writing only one block at a time. This is repeated 100 times and the average rate is determined. The file is then synched and closed, and
6

reopened with the same test repeated. A set of at least three files is used during the tests and averages are taken. Testing for random read is done on similar lines as random write.
Using the weights mentioned above, the rates for random reads, random writes, and sequential writes, are averaged to yield a net number for IOs per second. Then a CPU whose tpmC rating is known is taken and is used as the reference tpmC. The IOs per second obtained through the program is then normalized by dividing with a constant to match the reference tpmC. The weights for reads and writes, and the normalization factor are frozen, and the TPC meter tool is used to calibrate all other CPUs.
SPECjAppServer2004 is a popular benchmark across hardware vendors. For creating a JOPS meter a J2EE environment would be required. But it has been observed during capacity planning that JOPS and tpmC are proportional. Thus the TPC meter can also output a JOPS rating of the CPU core. The JOPS meter essentially takes the 10 rate calculated by TPC meter and divides by a normalizing factor to match a reference JOPS. The normalization factor then becomes a constant to estimate JOPS of all other CPUs.
TPC meter as described has been demonstrated to calibrate CPUs types across all leading vendors to date, viz. IBM, HP, SUN, AMD and Intel. The tool takes about 5 to 20 seconds to run the tests on the systems to be calibrated.
7

Table 1 provides the ratings on the various types of CPUs and Operating Systems.

Sno. CPU Type OS tpmC (per core) estimated by TPC meter Published tpmC result
1. HP Itanium 1.6Ghz HPUX11.23 41,000 43,000
2. IBM Power 4 AIX 5.2 26,000 24,000
3. IBM Power 5 1.9Ghz AIX 5.3 52,000 53,000
4. Intel Xeon 3.2Ghz RHEL4 38,000 37,000
5. SUNUltraSPARCIV+1.5Ghz Solaris 10 18,000 -
It can be seen from the above table that the tpmC ratings estimated by TPCMeter are quite close to the ones published by the vendors.
8

Table 2 provides the estimated JOPS per core, and compares against the published JOPS.

Sno. CPU Type OS Estimated JOPS by JOPS meter Published JOPS
1. AMD 2.2Ghz, dual core, processor 275 Solaris 10 50 50
2. AMD 2.2 Ghz, dual core, processor 275 RHEL4 110 100
3. AMD 2.2Ghz, single core, processor 848 RHEL4 85 -
4. HP Itanium 1.5Ghz HPUX11.23 130 120
5. IBM Power 4 AIX 5.2 85 -
6. IBM Power 5 1.9Ghz AIX 5.3 170 160
7. IntelXeon3.2Ghz RHEL4 125 120
8. SUN UltraSPARC IV+1.5Ghz Solaris 10 55 55
As can be seen, the JOPS prediction is fairly accurate.
9

Advantages:
The results obtained for tpmC and JOPS rating are encouraging. The tools can be used in a variety of ways including:
• Estimation of real tpmC and JOPS ratings during proposals. Even if a particular server is not available, the tool can be sent (in encrypted form) to a vendor lab and the output can be used for capacity planning purposes.
• Estimation of tpmC and JOPS of an existing server in production. If we expect workload to increase in the future and need to plan on more capacity with same or different types of CPUs, we can first determine the tpmC and JOPS of the CPU subsystem and then arrive at the desired tpmC and JOPS to support the additional workload.
• These two tools can be used to compare CPU performance for OLTP workloads. Rather than elaborate tests to determine which CPU will work out best for one's applications, a quick ballpark check can be done (even if required by adjusting the weights in these tools), to determine which CPU is likely to score higher.
While considerable emphasis has been placed herein on the particular features of TPC and JOPS meters, the improvisation with regards to it, it will be appreciated that various modifications can be made, and that many changes can be made in the preferred embodiment without departing from the principles of the invention. These and other modifications in the nature of the invention or the preferred embodiments will be apparent to those skilled in the art from the disclosure herein, whereby it is to be distinctly
10

understood that the foregoing descriptive matter is to be interpreted merely
as illustrative of the invention and not as a limitation.

Dated this 7th day of September 2007.

Mohan Dewan
Of R. K. Dewan & Co.
Applicants' Patent Attorney.

11