Inter-node messaging latency

[Yuji Ito]

Hi,

I'm investigating LWT performance with C* 3.11.3.
It looks that the performance is bounded by messaging latency when many
requests are issued concurrently.

According to the source code, the number of messaging threads per node is
only 1 thread for incoming and 1 thread for outbound "small" message to
another node.

I guess these threads are frequently interrupted because many threads are
executed when many requests are issued.
Especially, I think it affects the LWT performance when many LWT requests
which need lots of inter-node messaging are issued.

I measured that latency. It took 2.5 ms in average to enqueue a message at
a node and to receive the message at the **same** node with 96 concurrent
LWT writes.
Is it normal? I think it is too big latency, though a message was sent to
the same node.

Decreasing numbers of other threads like `concurrent_counter_writes`,
`concurrent_materialized_view_writes` reduced a bit the latency.
Can I change any other parameter to reduce the latency?
I've tried using message coalescing, but they didn't reduce that.

* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge

* C* configuration
- Apache Cassandra 3.11.3
- commitlog_sync: batch
- concurrent_reads: 32 (default)
- concurrent_writes: 32 (default)

Thanks,
Yuji

[sankalp kohli]

Inter-node messaging is rewritten using Netty in 4.0. It will be better to
test it using that as potential changes will mostly land on top of that.

Post by Yuji Ito
Hi,
I'm investigating LWT performance with C* 3.11.3.
It looks that the performance is bounded by messaging latency when many
requests are issued concurrently.
According to the source code, the number of messaging threads per node is
only 1 thread for incoming and 1 thread for outbound "small" message to
another node.
I guess these threads are frequently interrupted because many threads are
executed when many requests are issued.
Especially, I think it affects the LWT performance when many LWT requests
which need lots of inter-node messaging are issued.
I measured that latency. It took 2.5 ms in average to enqueue a message at
a node and to receive the message at the **same** node with 96 concurrent
LWT writes.
Is it normal? I think it is too big latency, though a message was sent to
the same node.
Decreasing numbers of other threads like `concurrent_counter_writes`,
`concurrent_materialized_view_writes` reduced a bit the latency.
Can I change any other parameter to reduce the latency?
I've tried using message coalescing, but they didn't reduce that.
* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge
* C* configuration
- Apache Cassandra 3.11.3
- commitlog_sync: batch
- concurrent_reads: 32 (default)
- concurrent_writes: 32 (default)
Thanks,
Yuji
---------------------------------------------------------------------

[Yuji Ito]

Hi,

Thank you for the reply.
I've measured LWT throughput in 4.0.

I used the cassandra-stress tool to insert rows with LWT for 3 minutes on
i3.xlarge and i3.4xlarge
For 3.11, I modified the tool to support LWT.
Before each measurement, I cleaned up all Cassandra data.

The throughput in 4.0 is 5 % faster than 3.11.
The CPU load of i3.4xlarge (16 vCPUs) is only up to 75% in both versions.
And, the throughput was slower than 4 times that of i3.xlarge.
I think the throughput wasn't bounded by CPU also in 4.0.

The CPU load of i3.4xlarge is up to 80 % with non-LWT write.

I wonder what is the bottleneck for writes on a many-core machine if the
issue about messaging has been resolved in 4.0.
Can I use up CPU to insert rows by changing any parameter?

# LWT insert
* Cassandra 3.11.3
| instance type | # of threads | concurrent_writes | Throughput [op/s] |
| i3.xlarge | 64 | 32 | 2815 |
| i3.4xlarge | 256 | 128 | 9506 |
| i3.4xlarge | 512 | 256 | 10540 |

* Cassandra 4.0 (trunk)
| instance type | # of threads | concurrent_writes | Throughput [op/s] |
| i3.xlarge | 64 | 32 | 2951 |
| i3.4xlarge | 256 | 128 | 9816 |
| i3.4xlarge | 512 | 256 | 11055 |

* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge / i3.4xlarge

* C* configuration
- Apache Cassandra 3.11.3 / 4.0 (trunk)
- commitlog_sync: batch
- concurrent_writes: 32, 256
- native_transport_max_threads: 128(default), 256 (when concurrent_writes
is 256)

Thanks,
Yuji

Post by sankalp kohli
Inter-node messaging is rewritten using Netty in 4.0. It will be better to
test it using that as potential changes will mostly land on top of that.

Post by Yuji Ito
Hi,
I'm investigating LWT performance with C* 3.11.3.
It looks that the performance is bounded by messaging latency when many
requests are issued concurrently.
According to the source code, the number of messaging threads per node is
only 1 thread for incoming and 1 thread for outbound "small" message to
another node.
I guess these threads are frequently interrupted because many threads are
executed when many requests are issued.
Especially, I think it affects the LWT performance when many LWT requests
which need lots of inter-node messaging are issued.
I measured that latency. It took 2.5 ms in average to enqueue a message
at a node and to receive the message at the **same** node with 96
concurrent LWT writes.
Is it normal? I think it is too big latency, though a message was sent to
the same node.
Decreasing numbers of other threads like `concurrent_counter_writes`,
`concurrent_materialized_view_writes` reduced a bit the latency.
Can I change any other parameter to reduce the latency?
I've tried using message coalescing, but they didn't reduce that.
* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge
* C* configuration
- Apache Cassandra 3.11.3
- commitlog_sync: batch
- concurrent_reads: 32 (default)
- concurrent_writes: 32 (default)
Thanks,
Yuji
---------------------------------------------------------------------

[Jeff Jirsa]

Are you sure you’re blocked on internode and not commitlog? Batch is typically not what people expect (group commitlog in 4.0 is probably closer to what you think batch does).
--
Jeff Jirsa

Post by Yuji Ito
Hi,
Thank you for the reply.
I've measured LWT throughput in 4.0.
I used the cassandra-stress tool to insert rows with LWT for 3 minutes on i3.xlarge and i3.4xlarge
For 3.11, I modified the tool to support LWT.
Before each measurement, I cleaned up all Cassandra data.
The throughput in 4.0 is 5 % faster than 3.11.
The CPU load of i3.4xlarge (16 vCPUs) is only up to 75% in both versions.
And, the throughput was slower than 4 times that of i3.xlarge.
I think the throughput wasn't bounded by CPU also in 4.0.
The CPU load of i3.4xlarge is up to 80 % with non-LWT write.
I wonder what is the bottleneck for writes on a many-core machine if the issue about messaging has been resolved in 4.0.
Can I use up CPU to insert rows by changing any parameter?
# LWT insert
* Cassandra 3.11.3
| instance type | # of threads | concurrent_writes | Throughput [op/s] |
| i3.xlarge | 64 | 32 | 2815 |
| i3.4xlarge | 256 | 128 | 9506 |
| i3.4xlarge | 512 | 256 | 10540 |
* Cassandra 4.0 (trunk)
| instance type | # of threads | concurrent_writes | Throughput [op/s] |
| i3.xlarge | 64 | 32 | 2951 |
| i3.4xlarge | 256 | 128 | 9816 |
| i3.4xlarge | 512 | 256 | 11055 |
* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge / i3.4xlarge
* C* configuration
- Apache Cassandra 3.11.3 / 4.0 (trunk)
- commitlog_sync: batch
- concurrent_writes: 32, 256
- native_transport_max_threads: 128(default), 256 (when concurrent_writes is 256)
Thanks,
Yuji

Inter-node messaging is rewritten using Netty in 4.0. It will be better to test it using that as potential changes will mostly land on top of that.

Post by Yuji Ito
Hi,
I'm investigating LWT performance with C* 3.11.3.
It looks that the performance is bounded by messaging latency when many requests are issued concurrently.
According to the source code, the number of messaging threads per node is only 1 thread for incoming and 1 thread for outbound "small" message to another node.
I guess these threads are frequently interrupted because many threads are executed when many requests are issued.
Especially, I think it affects the LWT performance when many LWT requests which need lots of inter-node messaging are issued.
I measured that latency. It took 2.5 ms in average to enqueue a message at a node and to receive the message at the **same** node with 96 concurrent LWT writes.
Is it normal? I think it is too big latency, though a message was sent to the same node.
Decreasing numbers of other threads like `concurrent_counter_writes`, `concurrent_materialized_view_writes` reduced a bit the latency.
Can I change any other parameter to reduce the latency?
I've tried using message coalescing, but they didn't reduce that.
* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge
* C* configuration
- Apache Cassandra 3.11.3
- commitlog_sync: batch
- concurrent_reads: 32 (default)
- concurrent_writes: 32 (default)
Thanks,
Yuji
---------------------------------------------------------------------

[Yuji Ito]

Hi Jeff,

I've not looked at the new inter-node latency in 4.0 yet.

I think it isn't blocked by commitlog.
In 3.11.3, I've probed each Paxos phase and commitlog sync.
(In the investigation, I didn't use cassandra-stress tool. The workload has
LWT read requests.)
The below table shows the average latency of each phase.
They are including inter-node messaging because I added `metrics` to
StorageProxy#cas().

It takes only 2,607 microseconds on average to sync commitlog in
`BatchCommitlogService`.
But each Paxos phase takes more than a few milliseconds except for the
commitlog sync.
Especially, though the read phase doesn't have write process, it takes
about 5 milliseconds.

MetricsLatency [us]
CAS Read 13556
CAS Write 32625
Prepare phase 8677
Read phase 4889
Propose phase 8706
Commit phase 10619

Thanks,
Yuji

Post by Jeff Jirsa
Are you sure you’re blocked on internode and not commitlog? Batch is
typically not what people expect (group commitlog in 4.0 is probably closer
to what you think batch does).
--
Jeff Jirsa
Hi,
Thank you for the reply.
I've measured LWT throughput in 4.0.
I used the cassandra-stress tool to insert rows with LWT for 3 minutes on
i3.xlarge and i3.4xlarge
For 3.11, I modified the tool to support LWT.
Before each measurement, I cleaned up all Cassandra data.
The throughput in 4.0 is 5 % faster than 3.11.
The CPU load of i3.4xlarge (16 vCPUs) is only up to 75% in both versions.
And, the throughput was slower than 4 times that of i3.xlarge.
I think the throughput wasn't bounded by CPU also in 4.0.
The CPU load of i3.4xlarge is up to 80 % with non-LWT write.
I wonder what is the bottleneck for writes on a many-core machine if the
issue about messaging has been resolved in 4.0.
Can I use up CPU to insert rows by changing any parameter?
# LWT insert
* Cassandra 3.11.3
| instance type | # of threads | concurrent_writes | Throughput [op/s] |
| i3.xlarge | 64 | 32 | 2815 |
| i3.4xlarge | 256 | 128 | 9506 |
| i3.4xlarge | 512 | 256 | 10540 |
* Cassandra 4.0 (trunk)
| instance type | # of threads | concurrent_writes | Throughput [op/s] |
| i3.xlarge | 64 | 32 | 2951 |
| i3.4xlarge | 256 | 128 | 9816 |
| i3.4xlarge | 512 | 256 | 11055 |
* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge / i3.4xlarge
* C* configuration
- Apache Cassandra 3.11.3 / 4.0 (trunk)
- commitlog_sync: batch
- concurrent_writes: 32, 256
- native_transport_max_threads: 128(default), 256 (when concurrent_writes is 256)
Thanks,
Yuji

Post by sankalp kohli
Inter-node messaging is rewritten using Netty in 4.0. It will be better
to test it using that as potential changes will mostly land on top of that.

Post by Yuji Ito
Hi,
I'm investigating LWT performance with C* 3.11.3.
It looks that the performance is bounded by messaging latency when many
requests are issued concurrently.
According to the source code, the number of messaging threads per node
is only 1 thread for incoming and 1 thread for outbound "small" message to
another node.
I guess these threads are frequently interrupted because many threads
are executed when many requests are issued.
Especially, I think it affects the LWT performance when many LWT
requests which need lots of inter-node messaging are issued.
I measured that latency. It took 2.5 ms in average to enqueue a message
at a node and to receive the message at the **same** node with 96
concurrent LWT writes.
Is it normal? I think it is too big latency, though a message was sent
to the same node.
Decreasing numbers of other threads like `concurrent_counter_writes`,
`concurrent_materialized_view_writes` reduced a bit the latency.
Can I change any other parameter to reduce the latency?
I've tried using message coalescing, but they didn't reduce that.
* Environment
- 3 node cluster
- Replication factor: 3
- Node instance: AWS EC2 i3.xlarge
* C* configuration
- Apache Cassandra 3.11.3
- commitlog_sync: batch
- concurrent_reads: 32 (default)
- concurrent_writes: 32 (default)
Thanks,
Yuji
---------------------------------------------------------------------