This is an interesting topic. I'd like to discuss this a bit during today's ODP public call.
I think the issue is that while a ring is a very useful implementation construct its semantics are very SW centric. Perhaps there's opportunity here for a new Queue type that would permit an implementation to implement the queue API as a ring for additional performance? The scheduler itself could use this since its use of queues is subject to the same issues. On Mon, Dec 7, 2015 at 11:39 PM, HePeng <xnhp0...@icloud.com> wrote: > Hi Maxim, > I implement a new version of cuckoo hash based on the ODP > buffer/pool. > > As I’ve mentioned earlier, the use of ring in cuckoo hash is like > to the use of > a container, e.g. a queue in C++ STL. In current ODP implementation, I > found that > the ODP queue is implemented more likely a facility for transmitting > objects, not > a container to store objects. Look at the ODP queue enqueue interface: > > int odp_queue_enq(odp_queue_t queue, odp_event_t ev); > > the *odp_event_t* is related to the events, but I only want to use > odp_queue to > storing and retrieving objects, any objects. > > > So I use ODP buffer/pool interfaces instead of ODP queue > for the new cuckoo hash implementation. > > However, compared to the previous implementation based on ring, > this version > suffers a serious performance degradation. The evaluation is carried out > on a Intel > servers. I test lookup time for 1000 lookups on a table storing 1 million > items. > The ODP buffer/pool version suffers at least a 2x performance degradation. > > This is the buffer/pool version, for 1M insert, and 1000 lookup time: > > Average insert time = 2.383836, lookup time = 0.000353, > > This is the ring version. > > Average insert time = 1.629115, lookup time = 0.000098 > > This performance degradation stems from the heavy implementation of > ODP buffer/pool. In the ring based one, all the key is stored in a big > array, and > the ring only stores the array indexes of each key. Keys are retrieved > using array indexes. > In the new one, I use ODP buffer to store the key content. Keys are > retrieved by > dereferencing a *odp_buffer_t* handle. > > With this high performance degradation, I suggest moving ring into > the odp/api > as a container implementation, and use the previous implementation of > cuckoo hash. > > > > > > > _______________________________________________ > lng-odp mailing list > lng-odp@lists.linaro.org > https://lists.linaro.org/mailman/listinfo/lng-odp >
_______________________________________________ lng-odp mailing list lng-odp@lists.linaro.org https://lists.linaro.org/mailman/listinfo/lng-odp
