Thanks for the suggestion. I was able to get better results tuning the GC settings but still not that great. I was seeing reading the netflix blog for the settings they have done and they have posted on blog. But i could not get close to what they are saying.

On Thu, Jul 19, 2012 at 9:45 PM, aaron morton <> wrote:
Three node cluster with replication factor of 3 gets me around 10 ms 100% writes with consistency equal to ONE. The reads are really bad and they are around 65ms. 
Using CL ONE in that situation, with a test that runs in a tight loop, can result in the clients overloading the cluster. 

Every node is a replica, so a write at CL ONE only has to wait for the local not to ACK. It will then return to the client before the remote nodes ACK, which means the client can send another request very quickly. In normal operation this may not be an issue, but load tests that run in a tight loop do not generate normal traffic. 

A better approach is to work at QUOURM so that network latency slows down individual client threads. Or generating the traffic using the Poisson distribution. The new load test from twitter uses that or you can use numpy for python. 


Aaron Morton
Freelance Developer

On 18/07/2012, at 11:29 PM, Manoj Mainali wrote:

How kind of client are you using in YCSB? If you want to improve latency, try distributing the requests among nodes instead of stressing a single node, try host connection pooling instead of creating connection for each request. Check high level clients like hector or asyantax for use if you are not already using them. Some clients have ring aware request handling.

You have a 3 nodes cluster and using a RF of three, that means all the node will get the data. What CL are you using for writes? Latency increases for strong CL. 

If you want to increase throughput, try increasing the number of clients. Of course, it doesnt mean that throughtput will always increase. My observation was that it will increase and after certain number of clients throughput decrease again.

Manoj Mainali

On Wednesday, July 18, 2012, Code Box wrote:
The cassandra stress tool gives me values around 2.5 milli seconds for writing. The problem with the Cassandra Stress Tool is that it just gives the average latency numbers and the average latency numbers that i am getting are comparable in some cases. It is the 95 percentile and 99 percentile numbers are the ones that are bad. So it means that the 95% of requests are really bad and the rest 5% are really good that makes the average go down. I want to make sure that the 95% and 99% values are in one digit milli seconds. I want them to be single digit because i have seen people getting those numbers. 

This is my conclusion till now with all the investigations:-

Three node cluster with replication factor of 3 gets me around 10 ms 100% writes with consistency equal to ONE. The reads are really bad and they are around 65ms. 

I thought that network is the issue so i moved the client on a local machine. Client on the local machine with one node cluster gives me again good average write latencies but the 99%ile and 95%ile are bad. I am getting around 10 ms for write and 25 ms for read. 

Network Bandwidth between the client and server is 1 Gigabit/second. I was able to at the max generate 25 K requests. So it could be the client is the bottleneck. I am using YCSB. May be i should change my client to some other.

Throughput that i got from a client at the maximum local was 35K and remote was 17K.

I can try these things now:-

Use a different client and see how much numbers i get for 99% and 95%. I am not sure if there is any client that gives me this detailed or i have to write one of my own.

Tweak some hard disk settings raid0 and xfs / ext4 and see if that helps.

Could be a possibility that the cassandra 0.8 to 1.1 the 95% and 99% numbers have gone down.  The throughput numbers have also gone down.

Is there any other client that i can use except the cassandra stress tool and YCSB  and what ever numbers i have got are they good ?

--Akshat Vig.

On Tue, Jul 17, 2012 at 9:22 PM, aaron morton <> wrote:
I would benchmark a default installation, then start tweaking. That way you can see if your changes result in improvements.
To simplify things further try using the tools/stress utility in the cassandra source distribution first. It's pretty simple to use. 

Add clients until you see the latency increase and tasks start to back up in nodetool tpstats. If you see it report dropped messages it is over loaded.

Hope that helps.

Aaron Morton
Freelance Developer

On 18/07/2012, at 4:48 AM, Code Box wrote:

Thanks a lot for your reply guys. I was trying fsyn = batch and window =0ms to see if the disk utilization is happening full on my drive. I checked the  numbers using iostat the numbers were around 60% and the CPU usage was also not too high. 

Configuration of my Setup :-

I have three m1.xlarge hosts each having 15 GB RAM and 4 CPU. It has 8 EC2 Compute Units.
I have kept the replication factor equal to 3. The typical write size is 1 KB. 

I tried adding different nodes each with 200 threads and the throughput got split into two. If i do it from a single host with FSync Set to Periodic and Window Size equal to 1000ms and using two nodes i am getting these numbers :-

[OVERALL], Throughput(ops/sec), 4771
[INSERT], AverageLatency(us), 18747
[INSERT], MinLatency(us), 1470
[INSERT], MaxLatency(us), 446413
[INSERT], 95thPercentileLatency(ms), 55
[INSERT], 99thPercentileLatency(ms), 167

[OVERALL], Throughput(ops/sec), 4678
[INSERT], AverageLatency(us), 22015
[INSERT], MinLatency(us), 1439
[INSERT], MaxLatency(us), 466149
[INSERT], 95thPercentileLatency(ms), 62
[INSERT], 99thPercentileLatency(ms), 171

Is there something i am doing wrong in cassandra Setup ?? What is the bet Setup for Cassandra to get high throughput and good write latency numbers ?

On Tue, Jul 17, 2012 at 7:02 AM, Sylvain Lebresne <>