5 ms·

Can anyone explain the Redis numbers, specifically the 3900X versus 5800X [1]? The 5800X is showing numbers easily more than 2x the 3900X. I understand this i

by bhauer 6y ago

Can anyone explain the Redis numbers, specifically the 3900X versus 5800X [1]? The 5800X is showing numbers easily more than 2x the 3900X. I understand this is a single-threaded test, so the additional cores of the 3900X (12C/24T versus 8C/16T for the 5800X) do it no advantage. But I wouldn't expect the Zen 3 architecture to be that dramatically superior to Zen 2. [1] https://raw.githubusercontent.com/tuhdo/tuhdo.github.io/master/emacs-tutor/static/5800x_m1_charts/Redis.png https://raw.githubusercontent.com/tuhdo/tuhdo.github.io/mast...

ben-schaaf 6y ago
This is pure speculation, but the Zen 3 cores have access to double the cache. In very specific workloads that could lead to a 2x speedup.
- stickyricky 6y ago
  I agree. I also think the unified core complex means less cross ccx cache access routes.
injinj 6y ago
I would bet that the redis-benchmark and the redis-server were running across the CCD, not tasklocked to CPUs on the same chiplet. I have a 3970x, I'll run this test on different chiplets and the same chiplet...
- injinj 6y ago
  $ taskset -c 1 redis-server $ taskset -c 2 redis-benchmark -p 6379 -P 8 -q -c 1 -n 100000 PING_INLINE: 775193.81 requests per second PING_BULK: 869565.19 requests per second SET: 724637.69 requests per second GET: 787401.56 requests per second INCR: 775193.81 requests per second LPUSH: 684931.50 requests per second RPUSH: 699300.69 requests per second LPOP: 680272.12 requests per second RPOP: 704225.31 requests per second SADD: 769230.81 requests per second HSET: 675675.69 requests per second SPOP: 819672.12 requests per second LPUSH (needed to benchmark LRANGE): 680272.12 requests per second LRANGE_100 (first 100 elements): 107758.62 requests per second LRANGE_300 (first 300 elements): 30721.96 requests per second LRANGE_500 (first 450 elements): 20040.08 requests per second LRANGE_600 (first 600 elements): 14658.46 requests per second MSET (10 keys): 317460.31 requests per second $ taskset -c 20 redis-server $ taskset -c 2 redis-benchmark -p 6379 -P 8 -q -c 1 -n 100000 PING_INLINE: 469483.56 requests per second PING_BULK: 505050.50 requests per second SET: 446428.56 requests per second GET: 465116.28 requests per second INCR: 458715.59 requests per second LPUSH: 425531.91 requests per second RPUSH: 438596.50 requests per second LPOP: 429184.56 requests per second RPOP: 434782.59 requests per second SADD: 460829.50 requests per second HSET: 421940.94 requests per second SPOP: 478468.88 requests per second LPUSH (needed to benchmark LRANGE): 425531.91 requests per second LRANGE_100 (first 100 elements): 96899.23 requests per second LRANGE_300 (first 300 elements): 29019.15 requests per second LRANGE_500 (first 450 elements): 19109.50 requests per second LRANGE_600 (first 600 elements): 14035.09 requests per second MSET (10 keys): 253807.11 requests per second
rfoo 6y ago
As the sibling comment said, this is most likely a methodology error. The difference is due to the 3900X benchmark goes plainly wrong. I guess the take home lesson is if someone is benchmarking a Zen-like CPU without mentioning "NUMA" at least once in their post, just don't trust it.

by bhauer 6y ago

ben-schaaf 6y ago

This is pure speculation, but the Zen 3 cores have access to double the cache. In very specific workloads that could lead to a 2x speedup.

stickyricky 6y ago
I agree. I also think the unified core complex means less cross ccx cache access routes.

injinj 6y ago

I would bet that the redis-benchmark and the redis-server were running across the CCD, not tasklocked to CPUs on the same chiplet. I have a 3970x, I'll run this test on different chiplets and the same chiplet...

injinj 6y ago
$ taskset -c 1 redis-server $ taskset -c 2 redis-benchmark -p 6379 -P 8 -q -c 1 -n 100000 PING_INLINE: 775193.81 requests per second PING_BULK: 869565.19 requests per second SET: 724637.69 requests per second GET: 787401.56 requests per second INCR: 775193.81 requests per second LPUSH: 684931.50 requests per second RPUSH: 699300.69 requests per second LPOP: 680272.12 requests per second RPOP: 704225.31 requests per second SADD: 769230.81 requests per second HSET: 675675.69 requests per second SPOP: 819672.12 requests per second LPUSH (needed to benchmark LRANGE): 680272.12 requests per second LRANGE_100 (first 100 elements): 107758.62 requests per second LRANGE_300 (first 300 elements): 30721.96 requests per second LRANGE_500 (first 450 elements): 20040.08 requests per second LRANGE_600 (first 600 elements): 14658.46 requests per second MSET (10 keys): 317460.31 requests per second $ taskset -c 20 redis-server $ taskset -c 2 redis-benchmark -p 6379 -P 8 -q -c 1 -n 100000 PING_INLINE: 469483.56 requests per second PING_BULK: 505050.50 requests per second SET: 446428.56 requests per second GET: 465116.28 requests per second INCR: 458715.59 requests per second LPUSH: 425531.91 requests per second RPUSH: 438596.50 requests per second LPOP: 429184.56 requests per second RPOP: 434782.59 requests per second SADD: 460829.50 requests per second HSET: 421940.94 requests per second SPOP: 478468.88 requests per second LPUSH (needed to benchmark LRANGE): 425531.91 requests per second LRANGE_100 (first 100 elements): 96899.23 requests per second LRANGE_300 (first 300 elements): 29019.15 requests per second LRANGE_500 (first 450 elements): 19109.50 requests per second LRANGE_600 (first 600 elements): 14035.09 requests per second MSET (10 keys): 253807.11 requests per second

rfoo 6y ago

As the sibling comment said, this is most likely a methodology error. The difference is due to the 3900X benchmark goes plainly wrong. I guess the take home lesson is if someone is benchmarking a Zen-like CPU without mentioning "NUMA" at least once in their post, just don't trust it.