90 likes | 260 Views
Detours. On a time-sharing system, a process is interrupted at any time. time. Detour is time period not for a target process, which contains. kernel service scheduler,memory management,network subsystem,etc other process etc cache miss, TLB miss. Record each detour in precious. start1.
E N D
Detours On a time-sharing system, a process is interrupted at any time. time Detour is time period not for a target process, which contains • kernel service • scheduler,memory management,network subsystem,etc • other process • etc • cache miss, TLB miss
Record each detour in precious start1 start2 start3 start4 ... ... startN time duration1 duration2 duration3 duration3 durationN record detours detour recorder output a recorded data # HEADERS # TICKS_PER_SECOND # ... start1 duration1 start2 duration2 start3 duration3 .....
Detour Recording Logic idea [Basic Idea – meta code] rdtsc() is called frequently Some experimental results, Running code at BG/L IO node 700MHz PowerPC Linux 2.4.19 based kernel rdtsc() is called every 0.1usec Minimum detour activety is 1.8 usec, which is timer update. If array is 2000 in size(1000 detours), it can record approx. 10 sec. current = rdtsc(); while( array is not full ) { prev = current; current = rdtsc(); if( current – prev > precision ) { array.push_back(prev); array.push_back(current); } }
Graph - timeseries 60ms 10ms
Graph – sorted by detour duration 60ms periodic event 10ms periodic event
Detours in fixed time quantum time Detours Q Total detour in Q Sync Frequency = 1/Q (HZ)
A node in parallel machine has detour. Detour distribution is same as single machine. node1 node2 node3 nodeN-1 nodeN pick detour randomly from single node experimental. fixed time record changing sync. freq.
Simulation algorithm for parallel machine (from a single node data) node1 node2 node3 nodeN-1 nodeN max(D)-min(D) try 1 d1 d2 d3 dN-1 dN max(D)-min(D) try 1 d1 d2 d3 dN-1 dN max(D)-min(D) try M d1 d2 d3 dN-1 dN [Input] recorded data N=machine size M=try count F=Sync Frequency [Output] average and standard deviation of {max(D)-min(D)}
What we know so far • maximum detour(latency) is a matter of parallel perfomance • minimize the max latency increase the parallel performance • linux kernel may be good for parallel machine