Python Memory Benchmark

These are the results of a very basic memory benchmark (written in Python) on my laptop.

Laptop manufacturer specs:

Lenovo - T460
CPU i5-6200U: 2.3GHz (0.43 ns per cycle) (2 cores with hyperthreading)
L1 cache 64 KB, hit rate 4 cycles = 1.72 ns
L2 cache 512 KB, hit rate 10 cycles = 4.3 ns
L3 cache 3072 KB, hit rate 40 cycles = 17.2 ns
RAM DDR3-3200: 32GB, 1600 MHz (1.25 ns per cycle), CAS Latency 16 cycles, Total latency = CAS latency x cycle = 16*1.25 = 20 ns, Throughput 25.6 GB/s

Summary plot (details below)

Create on disk a collection of N numpy arrays of size S from 2^9 to 2^28 bytes for all powers of two in between, that is from 512B up to 256MB. The arrays are of data type float64. Given that each float64 item is 8-bytes, this means the arrays are of length L from 8 to 33,554,432.

Load the whole collection in memory by reading it in a single numpy array of size (N,L) (good) or (L,N) (bad). The corresponding loop is

# good loop
for idx in range(N):
    ts[:, idx] = np.load(file)

# bad loop
for idx in range(N):
    ts[idx, :] = np.load(file)

Time the bad and the good loop
Results files are named result_N_S for each N and S. Each result file contains (size, time bad (s), time good (s)).
The plot above shows the ratio between the bad and the good timing as a function of S for N=30.