> On Wed, 11 Jul 2018 20:53:46 +0300
> Alex Kiselev <a...@therouter.net> wrote:
>> librte_lpm: Improve lpm6 performance
...
>>
>> /* LPM Tables. */
>> - struct rte_lpm6_rule *rules_tbl; /**< LPM rules. */
>> + struct rte_mempool *rules_pool; /**< LPM rules mempool. */
>> + struct rte_hash *rules_tbl; /**< LPM rules. */
>> struct rte_lpm6_tbl_entry tbl24[RTE_LPM6_TBL24_NUM_ENTRIES]
>> __rte_cache_aligned; /**< LPM tbl24 table. */
>> struct rte_lpm6_tbl_entry tbl8[0]
>> @@ -93,22 +106,81 @@ struct rte_lpm6 {
>> * and set the rest to 0.
> What is the increased memory overhead of having a hash table?
compared to the current rules array it's about 2 times since a prefix is stored
in
a rule (mempool) and in a rule key (hashtable).
I am only talking here about the rule storage.
And I've just realised it doesn't have to be this
way, I don't need the rules mempool anymore. I only need the rules hashtable,
since
it could contains everything a rule needs. A rule prefix is stored in a hash
key,
and a next hop index could be stored in a hash value. That would eliminate
memory overhead.
I'll try this way in next patch series.
> Wouldn't it make more sense to use something like tree, and use left/right
> in the rules entry. That way the memory is spread and scales with the number
> of rules.
Maybe. But there is no tree library in the DPDK. So I choose
a fast and simple way to implement
rules storage using the existent hashtable lib.
And it gives good perfomance results.
Anyway, it's not a data plane, add/delete operations are
executed not very often, so it's not critical to find
the most efficient (in terms of memory consumption) way, a good one is ok.
> Remember on a internet router, it is not unusual to 2M or more rules.
> Also. Please run checkpatch shell script on your patches. For example, there
> should be blank line between declarations and code.
I have. It didn't give me any warnings.
--
Alex