[
https://issues.apache.org/jira/browse/CASSANDRA-11519?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15332607#comment-15332607
]
Rei Odaira commented on CASSANDRA-11519:
----------------------------------------
Hi,
The linked document says PowerPC does not support unaligned 64-bit
floating-point access in the hardware, but it also says PowerPC supports
hardware-based unaligned integer access. As far as I understand, all of the
unaligned accesses in Memory.java, MemoryUtil.java, and FastByteOperations.java
are for integers.
Actually, the linked document is quite old (written in 2005). The latest POWER8
processor, released two years ago, supports hardware-based unaligned access for
every combination of 32-bit/64-bit integer/floating-point numbers. In my
measurement, there is no measurable performance penalty in unaligned access,
compared with aligned access. The only exception is access that crosses a page
boundary, which is 10x slower than normal access, but it is also the case in
x86.
If you need to know the situation in POWER7 and older processors, I'll
investigate them. There are some other implementations of the POWER
architecture (e.g. for game consoles), but I believe POWER8, which is targeting
the server market, is the only POWER implementation that practically runs
Cassandra.
> Add support for IBM POWER
> -------------------------
>
> Key: CASSANDRA-11519
> URL: https://issues.apache.org/jira/browse/CASSANDRA-11519
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Environment: POWER architecture
> Reporter: Rei Odaira
> Assignee: Rei Odaira
> Priority: Minor
> Fix For: 2.1.x, 2.2.x, 3.0.x, 3.x
>
> Attachments: 11519-2.1.txt, 11519-3.0.txt
>
>
> Add support for the IBM POWER architecture (ppc, ppc64, and ppc64le) in
> org.apache.cassandra.utils.FastByteOperations,
> org.apache.cassandra.utils.memory.MemoryUtil, and
> org.apache.cassandra.io.util.Memory.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
