Hi, Yes, We save the gc logs and can tell that the gc do its job that is why the total latency is very high SOMETIMES. However, we can not explain this strange behaviour which only occurs when we run the topology with 1 core. The strange behaviour: we run the topology several times and approximately , 3 out of 10 tests, we got very bad throughput which is about 50% of the normal throughput even though it is the same topology and test. Thanks On 02 Jun 2014, at 15:13, Nathan Leung <[email protected]> wrote:
> Stop the world garbage collection maybe? Do you have gc logs enabled? > > On Jun 2, 2014 9:06 AM, "dzacu1a" <[email protected]> wrote: > Hi, > We are doing a benchmark test and limiting the numbers of core used by Storm. > The topology contains 1 spout and 6 bolts. The first bolt has the highest > load. > Our experiment : we run the topology on one single machine with 4 cores. We > start with 1 core and then increasing to 2, 3 and 4 cores. We measure the > latency with a timestamp when the spout emits the tuple and when the tuple > reaches the last bolt. > The result : the latency results obtained from tests with 2,3 and 4 cores > are valid (as we expected - the latency results vary from 10 ms to 20 ms) > however we got strange values from the tests with only 1 core that we can’t > explain. The latency results vary from 4 ms to 120 ms. > Does anyone have any ideas why. > Thanks >
