-
|
I followed the instructions on the CacheLib docs and run the Cachebench with the following command: It seems the benchmark works all right and I get the following results. However, I am confused with the and |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments 10 replies
-
|
The only thing I have changed about |
Beta Was this translation helpful? Give feedback.
-
The numOps is the number of ops expressed per thread. So the overall ops would be numOps x numThreads
NumKeys is the number of unique keys that are used to generate the working set for the cache. If you set numKeys = 1 and numOps = 1M and 32 threads, cachebench would generate 32M requests total across 32 threads by repeatedly using the same single key (since numKey = 1). cachebench uses the popularity distribution to pick a suitable key for each operation.
What is the hardware configuration where you are running this benchmark ? Is your system bound by either compute/storage iops or bandwidth ? It does seem very low. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the answering. I am running the benchmark on an Ubuntu 18.04 server with 4 Intel Xeon Gold 6240 processors. Each processor runs at 2.6GHz and has 18 physical cores with hyperthreading enabled, a 32 kB instruction cache, 32 kB data cache, 1 MB L2 cache, and 25 MB L3 cache. The server has 375 GB of DDR4 DRAM. |
Beta Was this translation helpful? Give feedback.
-
|
@skyelves can you tell what resource (cpu/memory, SSD bandwidth, SSD iops) the system is constrained by when cachebench is running ? |
Beta Was this translation helpful? Give feedback.
-
|
It seems there are only several threads (no more than 10) running simultaneously. At first, the memory cost of CacheBench increases continuously while there is no writes to the SSD. And memory consumption increasing stops about 32 GB memory, so there are many free memory left. That's when there are increasing writes to the SSD. The average SSD write bandwidth is about |
Beta Was this translation helpful? Give feedback.
-
Can you show the results of the cachebench on your servers? I am wondering how significant the gap is between my results and the normal results. |
Beta Was this translation helpful? Give feedback.
-
|
You need to remove: opDelayNs as this statment limits overall workload performance. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for your answer. I removed the statement and run the benchmark again. |
Beta Was this translation helpful? Give feedback.
-
|
When I look into the However, it's wield that the hit ratio is not sensitive to the |
Beta Was this translation helpful? Give feedback.
-
Is this with or without hybrid cache ? cacheSizeMB is the DRAM cache size when used with hybrid cache. If the workload has nvmCacheSizeMB (hybrid cache) configured and if you shrink cacheSizeMB while continuing to have nvmCacheSizeMB being same, the overall hit rate should not change much. But you should notice the dram hit rate drop. You can check that in the following break down in the stats.
If you were instead referring to the pure DRAM setup and using uniform_values_get_throughput_test.json, and if the hit ratio is stable, then it could be because yyou are exiting the benchmark through Ctlr-C at around the same time and the hit ratio is determined by the number of sets that populated the cache until then and not the cache size. If this is the case, try "enableLookAside": true and watch the throughput until the evictions begin. |
Beta Was this translation helpful? Give feedback.
-
|
I am working without hybrid cache. Actually, I am using |
Beta Was this translation helpful? Give feedback.
-
|
@skyelves the operations are not ordered as |
Beta Was this translation helpful? Give feedback.
-
|
That makes sense. Thanks very much! |
Beta Was this translation helpful? Give feedback.
When I look into the
sizes.json, I found many value sizes are about 0.5 MB. Maybe it's because the value size is too large in the above cachebench.When I run another benchmark with the following command
./opt/cachelib/bin/cachebench -json_test_config ./cachelib/cachebench/test_configs/throughput/uniform_values_get_throughput_test.json, the throughput seems relatively normal .However, it's wield that the hit ratio is not sensitive to the
cacheSizeMB.I set
"numOps" : 4000000,and"valSizeRange" : [1, 1000],. When I changed from"cacheSizeMB" : 1024,to"c…