[jira] [Commented] (CASSANDRA-16218) Simplify the almost duplicated code for calculating serialization size and serialization
[
https://issues.apache.org/jira/browse/CASSANDRA-16218?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17218634#comment-17218634
]
Yifan Cai commented on CASSANDRA-16218:
---
The benchmark result shows the new approach takes *14% more time* on
calculating the serialized size. Attached the benchmark. So.. it is probably
not worthy. We may instead just fix CASSANDRA-16103 and pay more attention when
reviewing changes in serialize() and serializedSize().
{code:java}
c-16218
[java] BenchmarkMode Cnt Score Error
Units
[java] SerializationSizeBench.getSerializeSize avgt6 2.290 ± 0.489
ns/op
trunk
[java] BenchmarkMode Cnt Score Error
Units
[java] SerializationSizeBench.getSerializeSize avgt6 2.003 ± 0.163
ns/op{code}
> Simplify the almost duplicated code for calculating serialization size and
> serialization
>
>
> Key: CASSANDRA-16218
> URL: https://issues.apache.org/jira/browse/CASSANDRA-16218
> Project: Cassandra
> Issue Type: Improvement
> Components: Messaging/Client, Messaging/Internode
>Reporter: Yifan Cai
>Assignee: Yifan Cai
>Priority: Normal
> Fix For: 4.0-beta
>
> Attachments: SerializationSizeBench.java
>
>
> The current pattern of counting the serialization size and the actual
> serialization in the codebase is error-prone and hard to maintain. Those 2
> code paths have almost the same code repeated.
>
> The pattern can be found in {{org.apache.cassandra.net.Message#Serializer}}
> and numerous locations that use {{org.apache.cassandra.transport.CBCodec}}.
>
> I am proposing a new approach that lets both methods share the same code
> path. The code basically looks like the below (in the case of
> {{org.apache.cassandra.net.Message#Serializer}}).
> {code:java}
> // A fake DataOutputPlus that simply increment the size when invoking write*
> methods
> public class SizeCountingDataOutput implements DataOutputPlus
> {
> private int size = 0;
> public int getSize()
> {
>return size;
> }
> public void write(int b)
> {
>size += 1;
> }
> public void write(byte[] b)
> {
>size += b.length;
> }
> ...
> }
> {code}
> Therefore, in the size calculation, we can supply the fake data output and
> call serialize to get the size.
> {code:java}
> private int serializedSize(Message message, int version)
> {
> SizeCountingDataOutput out = new SizeCountingDataOutput();
> try
> {
>serialize(message, out, version);
> }
> catch (IOException exception)
> {
>throw new AssertionError("It should not happen!", exception);
> }
> return out.getSize();
> // The original implementation
> // return version >= VERSION_40 ? serializedSizePost40(message, version) :
> serializedSizePre40(message, version);
> }
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
-
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
[jira] [Commented] (CASSANDRA-16218) Simplify the almost duplicated code for calculating serialization size and serialization
[
https://issues.apache.org/jira/browse/CASSANDRA-16218?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17218577#comment-17218577
]
David Capwell commented on CASSANDRA-16218:
---
Idea sounds good to me, would simplify serializer logic. I took a quick scan
and the only negative I can find is extra function calls in some paths, but the
common case had roughly equal.
> Simplify the almost duplicated code for calculating serialization size and
> serialization
>
>
> Key: CASSANDRA-16218
> URL: https://issues.apache.org/jira/browse/CASSANDRA-16218
> Project: Cassandra
> Issue Type: Improvement
> Components: Messaging/Client, Messaging/Internode
>Reporter: Yifan Cai
>Assignee: Yifan Cai
>Priority: Normal
> Fix For: 4.0-beta
>
>
> The current pattern of counting the serialization size and the actual
> serialization in the codebase is error-prone and hard to maintain. Those 2
> code paths have almost the same code repeated.
>
> The pattern can be found in {{org.apache.cassandra.net.Message#Serializer}}
> and numerous locations that use {{org.apache.cassandra.transport.CBCodec}}.
>
> I am proposing a new approach that lets both methods share the same code
> path. The code basically looks like the below (in the case of
> {{org.apache.cassandra.net.Message#Serializer}}).
> {code:java}
> // A fake DataOutputPlus that simply increment the size when invoking write*
> methods
> public class SizeCountingDataOutput implements DataOutputPlus
> {
> private int size = 0;
> public int getSize()
> {
>return size;
> }
> public void write(int b)
> {
>size += 1;
> }
> public void write(byte[] b)
> {
>size += b.length;
> }
> ...
> }
> {code}
> Therefore, in the size calculation, we can supply the fake data output and
> call serialize to get the size.
> {code:java}
> private int serializedSize(Message message, int version)
> {
> SizeCountingDataOutput out = new SizeCountingDataOutput();
> try
> {
>serialize(message, out, version);
> }
> catch (IOException exception)
> {
>throw new AssertionError("It should not happen!", exception);
> }
> return out.getSize();
> // The original implementation
> // return version >= VERSION_40 ? serializedSizePost40(message, version) :
> serializedSizePre40(message, version);
> }
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
-
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
