Trial key for Aerospike VS Scylla benchmark

Benchmarks SSD Benchmarks (using ACT)
1

I have found very interesting comparison Aerospike VS Cassandra. Aerospike looks very interesting however there are still some open questions. For example Scylla says they faster than Cassandra in 2-5 times. Another point is 64 bytes per key - it is a lot, so intresting to compare in ALL-FLASH mode.

So I would test Aerospike vs Scylla but unfortunaly by default trial version just for 1 node. But much more useful find out how to work PROD-like Strong Consistency mode. I wrote to info @ aerospike.com but have no answer.

Could you please give some contacts which can provide trial version for 4 nodes?

2

Hi Boris!

Here is the test:

There inserts in Aerospike in 3.6x faster. ​I’ve done a long run test with Scylla and Aeropike (1 node and all-flash mode) and found Scylla is faster after few hours:

So if you share the results it would be very nice

3

Could you provide your workload, key set, configuration files, environment(Linux version, Aerospike version, sever model or instance type), and Aerospike log file. That would help us evaluate any issues you are running into.

4

Please note that you are using an all flash configuration and not the configuration utilized in the paper. So comparing the paper to your benchmark is not apples to apples comparison.

5

@tibors you are absolutely right, it is different test because ALL-FLASH much more load SSD. But it makes all DBs in the same position related the memory consumtion. I mean I limited memory consumtion by 32 GB (and Aerospike and Scylla).

The AWS was prepared by ansible from here:

I just change the memory setting, that’s all.

My workload is: Command line: -db site.ycsb.db.AerospikeClient -s -threads 50 -P workloads/workload -p as.host=127.0.0.1 -p as.namespace=test -p recordcount=4000000000 -p operationcount=4000000000 -p insertstart=0 -p readproportion=0 -p updateproportion=0 -p insertproportion=1 -p batchsize=1 -p fieldcount=10 -load

By the way - info @ aerospike.com is really not answer. There is the way to get trial for 4-8 nodes? Because test for 1 node is really not relevant for real life (usual nodes much more then 1) and it the obstacte to make independent evaluation. Mayby you can help to find the way to get the trial for few nodes?

6

If you can supply the configuration file that was generated from the Aerospike Ansible script that would be great. You should be able to find the config file in the /etc/aerospike directory. If you have the Aerospike log from your test run that would also provide some help in reviewing your issues. I can forward your message to the sales team an they can work with you on the license.

7

Did you use the automated all flash configuration. How did come up with the sprigs value in the configuration?

8

If you can supply the configuration file

Yes, it was generated by scripts:

namespace test {

 replication-factor 1

 memory-size 32G

 default-ttl 0d

 nsup-period 600

 rack-id 100000

 storage-engine device {

  device /dev/nvme1n1p2

  device /dev/nvme2n1p2

  data-in-memory false

 }

 partition-tree-sprigs 2048

 index-type flash {

  mount /nvme_all_flash_0    

  mount /nvme_all_flash_1    

  mounts-size-limit 3490G
}}

I can forward your message to the sales team and they can work with you on the license.

Thanks! You would be very nice)

9

I was reviewing your configuration. Since you are configuring all flash, you should review the documentation for configuring all flash. I believe the Ansible documentation points you to the all flash documentation but here are the links.

https://docs.aerospike.com/docs/operations/plan/capacity/index.html#aerospike-all-flash https://docs.aerospike.com/docs/operations/configure/namespace/index/index.html#flash-index

It is important that you size and configure all flash correctly or else you will see performance problems. Based on the number of keys in your workload, the partition-tree-sprigs should be 32768. I also see that the nsup-period is 600. Since you have a default-ttl of 0 you can set this to a high number to minimize it starting the nsup when it is not needed. Your mount size limit seems high. You should review the documentation to get that sized correctly.

What AWS instance type are you using?

10

Thanks, now it is much better - grow up from ~45K to 105K (i3en.x6large)

11

@tibors I didn’t get any feedback about the trial key. Is it possible to get it for testing in StrongConsistency mode?

12

You can get an Aerospike Database Enterprise Edition evaluation license for a single node EE cluster, with access to all the features. Not sure if that’s useful for you testing SC mode - you need a multi-node cluster for that to be meaningful. Also, be aware that the evaluation license agreement states you can’t publish benchmarks, it’s for your personal evaluation.

If you read Aerospike 4.0, Strong Consistency, and Jepsen you’ll see that in session consistency, two replicas (RF=2), and a stable cluster, there’s no performance difference from AP (‘High Availability’) mode.