We may reimplement such changes in the near feature considering also the new
status of DRAMSys (more specifically Dram.h) and DRAMPower.
Changes introduced in the following commits were reverted:
98249947f4c0f83bb1dc
Other changes:
Upper and lower limits for some plots adjusted.
Now we store the average power values directly obtained from the DRAMPower
library into the database. Thus we do not have to divide energy by time to
calculate them anymore (avoiding any possible lost of accuracy due to
rounding).
Other changes:
- At the very beginning (zero clock cycles) the energy is 0, so the
powerWindow thread wait first.
- When working with floats, we have to decide ourselves what is an acceptable
definition for "equal". Now we compare the energy value with a suitable
error margin (0.00001). Now the assertion that ensures the energy is not
zero is working properly.
Notes:
- The assert() function does nothing if NDEBUG is defined.
- The total energy is provided by DRAMPower as a double. We accumulate the
total energy for every window since we are clearing the library counters.
Explanation about double precision:
The double type ensures 15 decimal digits to represent a number. It does not
matter if the digits are before or after the comma. Thus we only have
rounding for numbers represented with more than 15 decimal digits.
In more technical terms:
An IEEE double has 53 significant bits (see also DBL_MANT_DIG in <cfloat>).
That is approximately 15.95 decimal digits (log10(2^53)). The implementation
sets the number of digits (DBL_DIG) to 15, not 16, because it has to round
down.
This way eventual changes to the default "sim-batch.xml" file will not affect
the result of this test.
Updated the expected values (expected.txt) after running the test with its own
sim-batch config file.
We used EXPLAIN QUERY PLAN and detected differences between machines.
Matthias found out that faster machines (DRAM server) use the "ranges_index"
for the "Transactions" table. But other machines (our laptops) use the primary
key defined for that table. Removing the primary key the performance issues
were solved.
