From aa6a6040e34c0c081a7359cee890e23bf7ff383e Mon Sep 17 00:00:00 2001 From: Grzegorz Hasse Date: Wed, 6 Dec 2017 16:32:27 -0800 Subject: Update documentation for `low_power_mode`. --- doc/tuning.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) (limited to 'doc/tuning.md') diff --git a/doc/tuning.md b/doc/tuning.md index 8eeefcb..bb9a7e4 100644 --- a/doc/tuning.md +++ b/doc/tuning.md @@ -9,6 +9,8 @@ * [Add more GPUs](#add-more-gpus) * [Increase Memory Pool](#increase-memory-pool) * [Scratchpad Indexing](#scratchpad-indexing) +* [CPU Backend](#cpu-backend) + * [Choose Value for `low_power_mode`](#choose-value-for-low_power_mode) ## NVIDIA Backend @@ -87,3 +89,15 @@ export GPU_SINGLE_ALLOC_PERCENT=99 The layout of the hash scratchpad memory can be changed for each GPU with the option `strided_index` in `amd.txt`. Try to change the value from the default `true` to `false`. + +## CPU Backend + +By default the CPU backend can be tuned in the config file `cpu.txt` + +### Choose Value for `low_power_mode` + +The optimal value for `low_power_mode` depends on the cache size of your CPU, and the number of threads. + +The `low_power_mode` can be set to a number between `1` to `5`. When set to a value `N` greater than `1`, this mode increases the single thread performance by `N` times, but also requires at least `2*N` MB of cache per thread. It can also be set to `false` or `true`. The value `false` is equivalent to `1`, and `true` is equivalent to `2`. + +This setting is particularly useful for CPUs with very large cache. For example the Intel Crystal Well Processors are equipped with 128MB L4 cache, enough to run 8 threads at an optimal `low_power_mode` value of `5`. -- cgit v1.1