hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andy Isaacson (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-3170) Add more useful metrics for write latency
Date Sat, 30 Jun 2012 01:05:44 GMT

    [ https://issues.apache.org/jira/browse/HDFS-3170?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13404350#comment-13404350
] 

Andy Isaacson commented on HDFS-3170:
-------------------------------------

{quote}
Can you use System.nanoTime so that it's (a) a monotonic clock instead of time-of-day, and
(b) higher granularity than MS? I think a lot of these metrics will end up sub-millisecond.
{quote}
You may want to review hdfs/server/common/Util.java:monotonicNow, although it returns milliseconds
rather than ns.
{quote}
They're not super expensive, but they are syscalls, so let's be efficient and call it the
minimal number of times.
{quote}
{{clock_gettime}} is a vsyscall, so it's really pretty cheap.  Looks like about 60 ns or 180
clock cycles for CLOCK_MONOTONIC (tested on Xeon X5670, 3GHz).  (Interestingly CLOCK_MONOTONIC_RAW
which I expected to be faster by skipping the adjustment, is actually slower at 240 ns on
the same Xeon.)

                
> Add more useful metrics for write latency
> -----------------------------------------
>
>                 Key: HDFS-3170
>                 URL: https://issues.apache.org/jira/browse/HDFS-3170
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>          Components: data-node
>    Affects Versions: 2.0.0-alpha
>            Reporter: Todd Lipcon
>            Assignee: Matthew Jacobs
>         Attachments: hdfs-3170.txt
>
>
> Currently, the only write-latency related metric we expose is the total amount of time
taken by opWriteBlock. This is practically useless, since (a) different blocks may be wildly
different sizes, and (b) if the writer is only generating data slowly, it will make a block
write take longer by no fault of the DN. I would like to propose two new metrics:
> 1) *flush-to-disk time*: count how long it takes for each call to flush an incoming packet
to disk (including the checksums). In most cases this will be close to 0, as it only flushes
to buffer cache, but if the backing block device enters congested writeback, it can take much
longer, which provides an interesting metric.
> 2) *round trip to downstream pipeline node*: track the round trip latency for the part
of the pipeline between the local node and its downstream neighbors. When we add a new packet
to the ack queue, save the current timestamp. When we receive an ack, update the metric based
on how long since we sent the original packet. This gives a metric of the total RTT through
the pipeline. If we also include this metric in the ack to upstream, we can subtract the amount
of time due to the later stages in the pipeline and have an accurate count of this particular
link.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Mime
View raw message