I try to follow https://igoro.com/archive/gallery-of-processor-cache-effects/ in python using numpy. Though it does not work and I don't quite understand why... numpy has fixed size dtypes, such as np.int64 which takes up 8 bytes. So with a cacheline of 64 bytes, 8 array values should be held in cache. Thus, when doing the timing, I should not see a notable change in required time when accessing values within the cacheline, because the same number of cache line transfers are needed. Based on this SO answer, I also tried to disable the garbage collection which didn't change anything. # import gc import time import numpy as np def update_kth_entries(arr, k): arr[k] = 0 start = time.perf_counter_ns() for idx in range(0, len(arr), k): arr[idx] = 0 end = time.perf_counter_ns() print(f"Updated every {k:4} th entry ({len(arr)//k:7} elements) in {(end - start)*1e-9:.5f}s") return arr # gc.disable() arr = np.arange(8*1024*1024, dtype=np.int64) print( f"(Data) size of array: {arr.nbytes/1024/1024:.2f} MiB " f"(based on {arr.dtype})" ) for k in np.power(2, np.arange(0,11)): update_kth_entries(arr, k) # gc.enable() This gives something like (Data) size of array: 64.00 MiB (based on int64) Updated every 1 th entry (8388608 elements) in 0.72061s Updated every 2 th entry (4194304 elements) in 0.32783s Updated every 4 th entry (2097152 elements) in 0.14810s Updated every 8 th entry (1048576 elements) in 0.07622s Updated every 16 th entry ( 524288 elements) in 0.04409s Updated every 32 th entry ( 262144 elements) in 0.01891s Updated every 64 th entry ( 131072 elements) in 0.00930s Updated every 128 th entry ( 65536 elements) in 0.00434s Updated every 256 th entry ( 32768 elements) in 0.00234s Updated every 512 th entry ( 16384 elements) in 0.00129s Updated every 1024 th entry ( 8192 elements) in 0.00057s Here is the output of lscpu -C NAME ONE-SIZE ALL-SIZE WAYS TYPE LEVEL SETS PHY-LINE COHERENCY-SIZE L1d 32K 384K 8 Data 1 64 1 64 L1i 32K 384K 8 Instruction 1 64 1 64 L2 256K 3M 4 Unified 2 1024 1 64 L3 16M 16M 16 Unified 3 16384 1 64 At this point I am quite confused about what I am observing. On the one hand I fail to see the cacheline using above code. On the other hand I can show some sort of CPU caching effect using something like in this answer with a large enough 2D array. I did above tests in a container on a Mac. A quick test on my Mac shows the same behavior. Is this odd behavior due to the python interpreter? What am I missing here? Continue reading...
