Measuring Performance Flashcards
Week 2.4 (38 cards)
what 3 things influence how many clock cycles instructions require on average
- instruction mix of given performance
- number of clock cycles per instruction
- processor architecture
what is the formula for cycles per instruction
CPI = sum of (CPIi x Ii) / Ic
where:
- Ii = executed instructions (type i)
- CPIi = cycles per instructions (type i)
- Ic = executed instructions (total)
- CPI = cycles per instructions (total)
what are the 2 formula for time taken to execute a program
T = Ic x CPI/f
OR
T = Ic(p x m.k)/f
where:
- p = cycles to decode & execute instructions
- m = number of memory references needed
- k = memory cycle time
- f = frequency
what does the ISA effect
Ic
p
what are the 4 main system attributes that affect performance factors
- ISA
- compiler technology
- processor implentation
- cache & memory hireachy
what performance factors does compiler technology affect
Ic - executed instructions
p - cycles to decode & execute instructions
m - number of memory references made
what performance factors does processor implementation affect
p
f - frequency
what performance factors does the cache & memory hierachy affect
k - memory cycle time
f
what are the 2 instructions per second formulae
IPSr = Ic/T
OR
IPSr = f/CPI
what is the formula for floating-point operations per second
FLOPSr = If/T
where:
- If = executed floating-point instructions (total)
what are the 6 basic measures of performance
- frequency
- CPI
- Instruction count
- Execution time
- IPS
- FLOPS
what is scaling
the ability of a system to adapt and handle increased workloads or user demands without performance degradation
what is the general scaling speed-up formula
Sp = T1/Tp
where:
- Tp = execution time with p processors
- p = number of processors
what is the general scaling parallel efficiency formula
Ep = Sp/p
what are the bounds of the general scaling speedup
1 <= Sp <= p
what are the bound of general scaling parallel efficency
1/p <= Ep <= 1
what are the 2 types of scaling
- strong
- weak
what law works with strong scaling
Amdahl’s Law
define amdahl’s law
the runtime of a parallel programme can be modelled as
Tp = s.T1 + (1 - s).T1/p where 0 <= s <= 1 id the fraction of code that must run serially
what are the 2 strong scaling parameters
- CONSTANT workload
- VARIABLE number of processors
what are the 3 strong scaling assumptions
- Serial and parallel fractions are easy to spot
- Parallel execution causes no overhead
- Workload can be divided in arbitrarily small chunks
what is the formula for speedup in strong scaling (with amdahl’s law)
Sp = T1/ Tp
= 1/s + (1 - s)/p
= p/ s.p + (1 - s)
what is the law that works with weak scaling
Gustafson’s Law
define Gustafson’s Law
the run time of a parallel programme follows T1 = s.Tp + (1 - s).Tp.p where 0 <= s <= 1 is the fraction of code that must run serially
