Name: Towards Energy Proportionality for Large-Scale Latency-Critical Workloads
1. Problem
1. Energy Proportionality for datacenter / OLDI
2. Challenges:
1. Idle power (not in PEGASUS paper)
2. Existing power management nor working properly (too aggressive & missing deadline OR not aggressive enough —> not proportional)
1. (inst / (v^2 * f * inst/s) < SLO )
2. User
1. operators of OLDI workloads (online (decline (SLO)) data intensive (sharing ))
uses lots of nodes so cant turn nodes off
3. Solution
1. Utilization <— queries/s / peak queries/s = load / capacity
for the particular system, max rate possible
2. SLO metrics: latency at “root node” —> query completion time (excludes internet) It makes sense since we do not sell internet but the datacenter product
1. Average over 30s latency of xx-percentile (i.g. 95, 99)
3. ISO-latency control
1. policy is top-down instead of bottom-up
2. each node has it own management so it can be turned off (but data intensive does not allow)
3. dedicated resource measure table +react
