Hi Rajiv, I was reading the paper "Enhancing the Scalability of Memcached" and found it interesting. nice work. Have you guys tested get on top of UDP? (I assume the published performance numbers are based on get over TCP).
Regards, Maysam UT Austin PhD student On Friday, June 29, 2012 11:00:00 AM UTC-7, rajiv wrote: > > Hi, > > > > For anyone interested, I just posted a new 1.6 branch called "bagLRU" at * > https://github.com/rajiv-kapoor/memcached/tree/bagLRU*<https://github.com/rajiv-kapoor/memcached/tree/bagLRU>that > includes an optimized engine called “bag_lru_engine”. The Bag LRU > engine implements striped locks for parallel hash table accesses and a > scheme for handling requests that give us lock-free GETs. Additionally I > have added a command line option to allow setting CPU affinity for > Memcached worker threads. I have noticed that CPU affinity along with NIC > queue (IRQ) affinity can help improve performance significantly. > > > > To use the Bag LRU engine use the “-E > path_to_memcached_root/engines/.libs/bag_lru_engine.so” option to Memcached > > To affinitize Memcached worker threads to “CPUs” (cores or logical > threads) use the command line option “-T n”. Where "n" is the increment > used for identifying the next CPU to bind the Memcached thread to. For > example, "-T 1" will bind CPUs sequentially to the Memcached threads > starting with CPU 0. “-T 2” will skip every other CPU starting with CPU > 0. Default is “-T 0” - meaning no thread affinity. > > > > Using mcblaster to generate a GETs only load, I have tested the Bag LRU > engine on a 16 core Sandybridge system from 1 to 16 threads and it shows > almost linear scaling with number of cores. For this test thread pinning > option used was “-T 1”. > > > > If you do try it out, I would love to get feedback on it. > > > > Thanks, > > \rajiv > Intel, Corp.
