experiment.md

Ember experiment schedule

Experiment list

• Spymemcached: the java client of memcached
• Spore: socc 2014
• RS: osdi 2016
• Ember

Parameters list

Client parameters

• client_type: 0 = MC, 2 = Ember/Spore, 3 = RS
• replica_mode: 0 = Ember, 1 = Spore

Server parameters

• is_detect_hot_spot: true | false
• replica_mode: 0 = Ember, 1 = Spore, 2 = RS, 3 = Others

Parameters table

Name	Client Parameters	Server Parameters
Spy	client_type = 0	is_detect_hot_spot = false
Spore	client_type = 2, replica_mode = 1	is_detect_hot_spot = true, replica_mode = 1
RS	client_type = 3	is_detect_hot_spot = false, replica_mode = 2
Ember	client_type = 2, replica_mode = 0	is_detect_hot_spot = true, replica_mode = 0

Inner parameters

Ember and Spore

• update_status_time: the period to get the workload of whole cluster
• hot_spot_buffer_size: size of the hot spot buffer, hot spots will be handled when the buffer is full
• hot_spot_manager_mode: the hotSpotManager mode, default mode is StreamHotSpotManager
• slice_time: the hotSpotManager period
• hot_spot_threshold: the hot spot threshold
• hot_spot_percentage: the hot spot percentage in the requests, such as 0.0001
• hot_spot_influence: the hot spot influence, such as 0.3
• bloom_filter_length: the length of bloom filter
• bloom_filter_number: the number of bloom filter
• frequent_percentage: the request for the frequent data, as 20/80, this parameter is 0.8

YcsbCore

• server_path: the path of server list
• data_hash_mode: data hash mode
• data_shards: data shard number
• parity_shards: parity shard number

Experiment Environment

• PC config: Ubuntu 14.04.3 LTS, 16 cores, 48 GB memory
• java version: 1.7.0_65
• data server: memcached 1.4.6
• workload generator: YCSB
• 2 data server (2 memcached instance each), 1 monitor server, 7 client

Experiment Result

The workloads are generated by YCSB, where each client can run 16 threads at most, so the most client threads is 7*16.

The throughput and 95%-latency relation.