Closed nagaradhesh closed 2 years ago
Hi, thank you for sharing us the experiment and result.
On your test condition, there are 2 SMDK contributions possibly related to the memcached performance.
Thanks for the response. "free page traversal : inter-zone traversal is cheaper than inter-node traversal." Can you please share more details on how page allocation is different b/w inter-zone and inter-node ? Please share if there are any links related to more information on the same.
Hello, nagaradhesh. Thank you for your interest. Please refer to the answer on a separate channel and let us know your thought.
We did two different memcached experiments in a CXL enabled host booted using SMDK kernel. For both, we ran memcached server on CXL exposed memory. First experiment, in default Zone partition setting, memcached server is run in exmem zone in compatible mode. Second experiment using numactl to redirect server traffic to CXL memory node1. We tried to plot Latency vs TPS(Transactions Per Second), we observed better latency numbers with Zone partition setting compared to numactl test for Node-1 CXL memory.
Can you please let us know if this is expected ? and is there any optimization in memory allocation for Zone partition. Following is the Configuration used: export CXLMALLOC_CONF=use_exmem:true,exmem_zone_size:12288,normal_zone_size:12288,maxmemory_policy:oom,use_auto_arena_scaling:false,priority:exmem