I am a newbie to Sniper.
I have an application that takes less than 3 seconds to complete on an Unbuntu 12 Linux box.
However it takes 227 minutes to complete under Sniper.
Is there a minimal configuration to get Sniper to run the fastest and producing consistent results?
Here is my current configuration:
#include base
[perf_model/cache]
levels = 3
[perf_model/l1_icache]
writethrough = 0
[perf_model/l1_dcache]
writethrough = 0
[perf_model/l2_cache]
writethrough = 0
[perf_model/l3_cache]
cache_block_size = 64
dvfs_domain = global
prefetcher = none
shared_cores = 4
writethrough = 0
perf_model_type = parallel
writeback_time = 0
tags_access_time = 10
data_access_time = 30 # 35 cycles total according to membench, +L1+L2 tag times
replacement_policy = lru
address_hash = mask
associativity = 16
cache_size = 8192
[perf_model/core/oneipc]
latency_cutoff = 4 # Maximum latency which is assumed to be completely overlapped. L1-D hit latency should be a good value
[perf_model/dram]
# -1 means that we have a number of distributed DRAM controllers (4 in this case)
num_controllers = -1
controllers_interleaving = 4
Your help in this matter will be greatly appreciated.
Zung