It sounds as if you are actually testing “vertical scalability” (load on a
single node) rather than Cassandra’s sweet spot of “horizontal scalability”
(add more nodes to handle higher load.) Maybe you could clarify your intentions
and specific use case.
Also, it sounds like you are trying to focus on large queries, but Cassandra’s
sweet spot is lots of smaller queries. With larger queries you can end up
measuring things like the capabilities of your hardware, cpu cores, memory, I/O
bandwidth, network latency, JVM configuration, etc. rather than measuring
Cassandra per se. So, again, maybe you could clarify your intended use case.
It might be that you need to add more “vertical scale” (bigger box, more cores,
more memory, beefier I/O and networking) to handle large queries, or maybe
simple, Cassandra-style “horizontal scaling” (adding nodes) will be sufficient.
Sure, you can tune Cassandra for single-node performance, but that seems lot a
lot of extra work, to me, compared to adding more cheap nodes.
-- Jack Krupansky
From: Diane Griffith
Sent: Thursday, July 17, 2014 9:31 AM
To: user
Subject: Re: trouble showing cluster scalability for read performance
Duncan,
Thanks for that feedback. I'll give a bit more info and then ask some more
questions.
Our Goal: Not to produce the fastest read but show horizontal scaling.
Test procedure:
* Inserted 54M rows where one third of that represents a unique key, 18M keys.
End result given our schema is the 54M rows becomes 72M rows in the column
family as the control query load to use.
* have a client that queries 100k records in configurable batches, set to 1k.
And then it does 100 reps of queries. It doesn't do the same keys for each
rep, it uses an offset and then it increases the keys to query.
* We can adjust the hit rate, i.e. how many of the keys will be found but have
been focused on 100% hit rate
* we run the query where multiple clients can be spawned to do the same query
cycle 100k keys but the offset is not different so each client will query the
same keys.
* We thought we should manually compact the tables down to 1 sstable on a given
node for consistent results across different cluster sizes
* We had set replication factor to 1 originally to not complicate things or
impact initial write times even. We would assess rf later was our thought.
Since we changed the keys getting queried it would have to hit additional nodes
to get row data but for just 1 client thread (to get simplest path to show
horizontal scaling, had a slight decrease of performance when going to 4 nodes
from 2 nodes)
Things seen off of given procedure and set up:
1.. 1 client thread: 2 nodes do better than 1 node on the query test. But 4
nodes did not do better than 2.
2.. 2 client threads: 2 nodes were still doing better than 1 node
3.. 10 client threads: the times drastically suffered and 2 nodes were doing
1/2 the speed of 1 node but before 1 to 2 threads performed better on 2 nodes
vs 1 node. There was a huge decrease in performance on 2 nodes and just a mild
decrease on 1 node.
Note: 50+ threads was also drastically falling apart.
Observations:
a.. compacting each node to 1 table did not seem to help as running 10 client
threads on exploded sstables and 2 nodes was 2x better than the last 2 node 10
client test but still decreased performance from 1 to 2 threads query against
compacted tables
b.. I would see upwards to 10 read requests pending at times while 8 to 10
were processing when I did nodetool tpstats.
c.. having key cache on or disabled did not seem to impact things noticeably
with our current configuration
.
Questions:
1.. can multiple threads read the same sstable at the same time? Does
compacting down to 1 sstable (to get a given row into one sstable) add any
benefit or actually hurt like limited testing has indicated currently?
2.. given the above testing process, does it still make sense to adjust
replication factor appropriately for cluster size (i.e. 1 for 1 node cluster, 2
for 2 node cluster, 3 for n size cluster). We assumed it was just the ability
for threads to connect into a coordinator that would help but sounds like it
can still block
I'm going to try a limited test with changing replication factor. But if
anyone has any input on compacting to 1 sstable benefit or detriment on just
simple scalability test, how if at all does cassandra block on reading
sstables, and if higher replication factors do indeed help produce reliable
results it would be appreciated. I know part of our charter was keep it simple
to produce the scalability proof but it does sound like replication factor is
hurting us if the delay between clients for the same keys is not long enough
given the fact we are not doing different offsets for each client thread.
Thanks,
Diane
On Thu, Jul 17, 2014 at 3:53 AM, Duncan Sands <duncan.sa...@gmail.com> wrote:
Hi Diane,
On 17/07/14 06:19, Diane Griffith wrote:
We have been struggling proving out linear read performance with our
cassandra
configuration, that it is horizontally scaling. Wondering if anyone has any
suggestions for what minimal configuration and approach to use to
demonstrate this.
We were trying to go for a simple set up, so on the keyspace and/or column
families we went with the following settings thinking it was the minimal to
prove scaling:
replication_factor set to 1,
a RF of 1 means that any particular bit of data exists on exactly one node.
So if you are testing read speed by reading the same data item again and again
as fast as you can, then all the reads will be coming from the same one node,
the one that has that data item on it. In this situation adding more nodes
won't help. Maybe this isn't exactly how you are testing read speed, but
perhaps you are doing something analogous? I suggest you explain how you are
measuring read speed exactly.
Ciao, Duncan.
SimpleStrategy,
default consistency level,
default compaction strategy (size tiered),
but compacted down to 1 sstable per cf on each node (versus using leveled
compaction for read performance)
*Read Performance Results:*
1 client thread - 2 nodes > 1 node was seen but we couldn't show increased
performance adding more nodes i.e 4 nodes ! > 2 nodes
2 client threads - 2 nodes > 1 node still was true but again we couldn't
show
increased performance adding more nodes i.e. 4 nodes ! > 2 nodes
10 client threads - this time 2 nodes < 1 node on performance numbers. 2
nodes
suffered from larger reduce throughput than 1 node was showing.
Where are we going wrong?
How have others shown horizontal scaling for reads?
Thanks,
Diane
