On 1/20/2013 2:13 PM, Markus Jelsma wrote:
Hi Shawn,

Although our heap spaces are much less than yours (256M for 2x 2.5GB cores per 
node) we saw decreased throughput and higher latency with G1 on Java 6. You can 
also expect higher CPU consumption. You can check it very well with VisualVM 
attached.

Looking forward to your results.

I don't have any really good test tools developed for testing throughput and latency. I have some less-than-ideal tools for other purposes that I might be able to adapt.

Throughput is not a major issue for us - query volume is quite low. I would be mildly surprised by 5 queries per second. I don't have much of an idea of queries per second over the short term - the numbers available in 3.5 are limited.

As for latency, early indications from an old SOLR-1972 patch suggest that the QTime values might be a little higher. The primary server stats (using CMS/ParNew) are over 1 million queries, and the secondary server stats (using G1) so far are only about 5000 queries. The QTime values are steadily dropping as the number of queries goes up.

Here's a status page that gathers all the stats. Chain A is using CMS/ParNew and is no longer receiving queries. All the queries are now going to chain B, which is using G1.

http://dl.dropbox.com/u/97770508/g1-vs-cms-stats.png

The server CPU utilization graph doesn't have enough information yet to make any determination, but what little data is visible suggests that CPU may be higher. The secondary servers also have slightly slower CPUs than the primary servers. I was forced to make concessions on later purchases to keep the cost down.

Thanks,
Shawn

Reply via email to