cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From maneela a <>
Subject Re: RackAwareStrategy vs RackUnAwareStrategy on AWS EC2 cloud
Date Fri, 09 Jul 2010 19:16:26 GMT
Thanks for your quick reply.. JoeI forgot to mentioned that we are using PropertyFileEndPointSnitch
to provide cassandra about our network topology and below is property file uses by that class
cat rack.properties10.9.0.6=east:r1b10.9.0.18=east:r1c10.9.0.14=east:r1d10.9.0.10=west:r1adefault=east:rdef
In my first glance, VPNCubed uses openVPN software as part of transport layer so I did not
look into it much deeper. I will work on it to see if it helps in out set up
we are looking for something like Oracle Primary/Standby solution where write operations happens
only on Primary set of nodes ( in our case nodes those are running in different AZ in East
region) and one copy of each data block should replicate to the node running in the West region
so that read operation can be available local to any region because we have applications that
consume cassandra data,  are running from both East and West regions. we are inclined to
accept write latency between regions because readers won't look for data immediately and there
will be around 5 -10 mins gap between write and read operations.
Is there any way to mark cassandra node to keep it as just for replication purpose and not
to be as Primary for any data range in the ring? 

--- On Fri, 7/9/10, Joe Stump <> wrote:

From: Joe Stump <>
Subject: Re: RackAwareStrategy vs RackUnAwareStrategy on AWS EC2 cloud
Date: Friday, July 9, 2010, 1:41 PM

We had similar issues when we started running Cassandra on EC2 between multiple AZ's (not
regions; we're working up to that shortly). We ended up building a rack aware strategy specific
to AWS, which is posted somewhere in JIRA. Basically it uses the AWS API to ensure that replicants
are stored in each AZ. We then ensure that our clients are only reading from nodes in a given
AZ. What I'm guessing is that you're seeing latency issues between regions combined with
a higher consistency level than what we use.
Also, SSL tunnels are hard to scale from the management side. The Amazon folks have told us
that VPNCubed is a better solution for such things.

On Jul 9, 2010, at 11:36 AM, maneela a wrote:
Are there any known performance issues if cassandra cluster launched with RackAwareStrategy
because I see huge performance difference between RackAwareStrategy vs RackUnAwareStrategy. 
Here are details:

we have a cluster setup with 4 EC2 X large nodes, 3 of them are running in East region and
4th one is running in West region and they all communicate with each other through VPN tunnel
interface which is only way we found to achieve ring architecture across Amazon cloud regions:

we are able to process 3.5K write operations per second when we used RackUnAwareStrategy whereas 
:/home/ubuntu/cassandra/contrib/py_stress# ./ -o insert -n 80000 -y regular -d
--threads 100 --keep-goingtotal,interval_op_rate,avg_latency,elapsed_time35935,3593,0.0289930914479,1070531,3459,0.0289145907593,2080000,946,0.0267288666213,30
whereas we are able to process only 250 write operations per second when we used RackAwareStrategy
:/home/ubuntu/cassandra/contrib/py_stress# ./ -o insert -n 80000 -y regular -d
--threads 100
Thanks in advance


View raw message