Hi Tim, On 03/01/2014 02:02 PM, Tim Wintle wrote:
137GB would fairly easily fit in core memory on a single node these days: so it seems a very low amount for a 27 node cluster..
Note that we only have 4 GB of RAM per node, so only 1 GB of Cassandra heap. Are you assuming large memory servers, or am I misunderstanding you?
Off the top of my head: would 99th percentile latency be improved by using replication factor 5, assuming you are doing quorum operations..
We are currently analyzing the case with reads/writes at consistency level ONE, so I don't think increasing the replication factor will help us right now.
I found this doc last night: http://software.intel.com/sites/default/files /Configuration_and_Deployment_Guide_for_Cassandra_on_IA.pdf The numbers that they quote for data size seem to be quite low as well. Thoughts? -Bill
Sent from my phone On 1 Mar 2014 14:33, "William Katsak" <wkat...@cs.rutgers.edu <mailto:wkat...@cs.rutgers.edu>> wrote: Hello, I am doing some academic work with Cassandra 1.1.6 (I am on an older version because of a bunch of implemented modifications that have been in the works for a while), and I am wondering if the list can help me resolve some questions I have. I am running a cluster of 27 nodes with the following configuration: Intel Atom (2 core) @ 1.8 GHz 4 GB RAM 250 GB HDD 64 GB SSD Gigabit Ethernet With this cluster size, I currently have loaded 135 GB of data (replicated * 3), giving me data of ~15 GB per node. I am using Leveled Compaction with a 5mb SSTable size. Commitlog is in HDD, data is on SSD. My workload is YCSB/uniform distribution/75% read-25% write. My questions are: - Is this a reasonable data size for this hardware? - What should be compaction throughput be set to? I am targeting a 99th percentile latency SLA, and it seems that compaction throughput greatly affects the 99th percentile latency. The guideline seems to be 16-32x insertion rate, but this slows down the 99th percentile time dramatically. In addition, there seems to be a feedback loop where if you insert faster, you need more compaction, but if you had more compaction, you can't insert as fast. What is best practice on this? - What is a reasonable operation throughput to expect from this configuration? Sorry for the info dump, but I have been fighting with this for a while now. I've tried to read everything I can about tuning and provisioning, but continue to have an issue where I can find a load rate that hits my 99th percentile SLA on average, but have large latency spikes that don't seem to match a pattern. Thanks in advance for any advice you can give, even if it is just "go read this document". Sincerely, Bill Katsak Ph.D. Student Rutgers University
