Re: [Gems-users] calculate network latency

[Guoqiang Yang <nkyangguoqiang@xxxxxxxxx>]

Hi Xu han,

You must have been using Garnet network model, which simulates a more realistic network model. In Garnet fixed router pipeline model
They assume the router latency itself is 4-cycle. That adds up to 10 cycles. You might want to look into the details of Garnet network
model in the following URL: http://www.cs.wisc.edu/gems/doc/gems-wiki/moin.cgi/Garnet

Hope that helps,

Guoqiang Yang
 
==========================
Hello, gentlemen,

May I ask a question about network latency in ruby:

when I use a topolody specified in file
'NUCA_Mesh_Procs-4_ProcsPerChip-4_L2Banks-4_Memories-4.txt', whose content
is:
processors:4
procs_per_chip:4
L2banks:4
memories:4
bw_unit:1000

ext_node:L1Cache:0 int_node:8 link_latency:1 bw_multiplier:32
ext_node:L1Cache:1 int_node:9 link_latency:1 bw_multiplier:32
ext_node:L1Cache:2 int_node:10 link_latency:1 bw_multiplier:32
ext_node:L1Cache:3 int_node:11 link_latency:1 bw_multiplier:32

ext_node:L2Cache:0 int_node:4 link_latency:20 bw_multiplier:32
ext_node:L2Cache:1 int_node:5 link_latency:20 bw_multiplier:32
ext_node:L2Cache:2 int_node:6 link_latency:20 bw_multiplier:32
ext_node:L2Cache:3 int_node:7 link_latency:20 bw_multiplier:32
ext_node:Directory:0 int_node:4 link_latency:100 bw_multiplier:16
ext_node:Directory:1 int_node:5 link_latency:100 bw_multiplier:16
ext_node:Directory:2 int_node:6 link_latency:100 bw_multiplier:16
ext_node:Directory:3 int_node:7 link_latency:100 bw_multiplier:16

int_node:0 int_node:1 link_latency:2 bw_multiplier:16
int_node:1 int_node:3 link_latency:2 bw_multiplier:16
int_node:2 int_node:0 link_latency:2 bw_multiplier:16
int_node:3 int_node:2 link_latency:2 bw_multiplier:16

int_node:4 int_node:8 link_latency:1 bw_multiplier:32
int_node:5 int_node:9 link_latency:1 bw_multiplier:32
int_node:6 int_node:10 link_latency:1 bw_multiplier:32
int_node:7 int_node:11 link_latency:1 bw_multiplier:32

int_node:8 int_node:0 link_latency:1 bw_multiplier:32
int_node:9 int_node:1 link_latency:1 bw_multiplier:32
int_node:10 int_node:2 link_latency:1 bw_multiplier:32
int_node:11 int_node:3 link_latency:1 bw_multiplier:32


THEN, in the result of topology print:
L1Cache-0 Network Latencies
 L1Cache-0 -> L2Cache-0 net_lat: 24
 L1Cache-0 -> L2Cache-1 net_lat: 31
 L1Cache-0 -> L2Cache-2 net_lat: 31
 L1Cache-0 -> L2Cache-3 net_lat: 34
 L1Cache-0 -> L1Cache-1 net_lat: 10
 L1Cache-0 -> L1Cache-2 net_lat: 10
 L1Cache-0 -> L1Cache-3 net_lat: 13
 L1Cache-0 -> Directory-0 net_lat: 104
 L1Cache-0 -> Directory-1 net_lat: 111
 L1Cache-0 -> Directory-2 net_lat: 111
 L1Cache-0 -> Directory-3 net_lat: 114
.......

I wonder, for example, how the smallest latency 'L1Cache-0 -> L1Cache-1
net_lat: 10' is caculated ?
According to the topology file, link latency between two nearby nodes is
2 and link latency to L1 cache is only 1. so should the latency between
two nearby L1 caches around 4 or 5?
but it says 10 in the result, what else shall we include?

Thanks a lot!

Xu Han