[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

20210903: Re: [conduit] 20210830: Re: High CONDUIT latencies from vm-lnx-conduit2.ncep.noaa.gov

Hi Pete, Kevin, et. al,

re:
> Unidata folks, are your files also incomplete due to the large lags, or
> is that specific to us due to a too-small queue? I am looking at the
> file sizes for the 0.25 deg runs on motherlode, and just looking at the
> file sizes, it appears Unidata also are missing data. Seems like a
> complete run should be ~47 Gb, and several are 41, 42, 39Gb..

We have all been experiencing increased latencies in the CONDUIT feed
over the past several days.  Here is a link to the CONDUIT latency
plot for the Unidata machine that REQUESTs CONDUIT via a 20-way
split from the NCEP top level CONDUIT cluster conduit.ncep.noaa.gov:

https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+emo.unidata.ucar.edu

As you will see, the CONDUIT latencies from the NCEP machine
vm-lnx-conduit1.ncep.noaa.gov have remained low throughout the time
covered by the plot while the latencies from vm-lnx-conduit2 have
risen back to previous levels after being significantly better
from about 22Z on 20210901 until about 22Z on 20210902.

As to whether or not the cause of missing fields in the data sets
that Pete is referring to are being caused by the high latencies,
I think that the answer is no.  My reasoning for saying no is the
LDM queues on all Unidata machines involved in the ingest and realy
of CONDUIT data are all large enough to hold > 70% of an hour if
not more, and the largest latencies being seen from
vm-lnx-conduit2.ncep.noaa.gov are less than 50% of an hour (1800 s).
My conclusion is that the products that are seen to be missing
are not being sent from NCEP.

re: Jesse Marks asked
> One question - do you know the specific day this problem first
> appeared?

The latency plot referred to above can be used to answer this question.

Cheers,

Tom

On 9/3/21 11:55 AM, Pete Pokrandt wrote:
Yes, we are seeing the large lags again, and losing data, I think because it ages out of our ldm product queue before we process it. Unidata folks, are your files also incomplete due to the large lags, or is that specific to us due to a too-small queue? I am looking at the file sizes for the 0.25 deg runs on motherlode, and just looking at the file sizes, it appears Unidata also are missing data. Seems like a complete run should be ~47 Gb, and several are 41, 42, 39Gb..
https://motherlode.ucar.edu/native/grid/NCEP/GFS/Global_0p25deg/ <https://motherlode.ucar.edu/native/grid/NCEP/GFS/Global_0p25deg/> 

GFS_Global_0p25deg_20210831_0000.grib2      2021-08-30 23:16   47G
GFS_Global_0p25deg_20210831_0600.grib2      2021-08-31 05:14   47G
GFS_Global_0p25deg_20210831_1200.grib2      2021-08-31 11:14   47G
GFS_Global_0p25deg_20210831_1800.grib2      2021-08-31 17:17   47G
GFS_Global_0p25deg_20210901_0000.grib2      2021-08-31 23:16   41G
GFS_Global_0p25deg_20210901_0600.grib2      2021-09-01 05:14   47G
GFS_Global_0p25deg_20210901_1200.grib2      2021-09-01 11:14   47G
GFS_Global_0p25deg_20210901_1800.grib2      2021-09-01 17:17   47G
GFS_Global_0p25deg_20210902_0000.grib2      2021-09-01 23:16   47G
GFS_Global_0p25deg_20210902_0600.grib2      2021-09-02 05:14   47G
GFS_Global_0p25deg_20210902_1200.grib2      2021-09-02 11:15   47G
GFS_Global_0p25deg_20210902_1800.grib2      2021-09-02 17:17   47G
GFS_Global_0p25deg_20210903_0000.grib2      2021-09-02 23:16   41G
GFS_Global_0p25deg_20210903_0600.grib2      2021-09-03 05:14   42G
GFS_Global_0p25deg_20210903_1200.grib2      2021-09-03 11:14   39G
If that is the case, that we are both dropping/losing data due to the lags, the maybe problem is that the product queues on the NCEP virtual machines are not large enough to handle the feed, does that seem correct?
My product queue on idd-agg is 78000 Mb (~78 Gb) - but that also handles other data such as the NEXRAD, NEXRAD2, HDS, etc.. however GOES16/17 data does NOT flow through that machine, so it doesn't contribute to the load there.
How big are the product queues on the Unidata CONDUIT ingest machines? Or on the NCEP source machines?  Could a possible solution to this be to artificially delay the ingest of the GFS grids a bit to lower the peak amount of data going through? Maybe sleep 20 seconds or a minute or whatever between ingesting each forecast hour? I'd personally rather get a complete data set a bit later than an incomplete data set on time.
Here's the graph of lag from conduit.ncep.noaa.gov to idd-agg.aos.wisc.edu from the past two days. When lags get > 1000 seconds or so, that's the point where we and our downstreams start dropping/losing data..
It would really be helpful to get this resolved, whether it is a fix for whatever's causing the large lags at NCEP, or us acquiring an ingest machine with enough memory to handle a larger product queue (if the issue is us and not the upstream product queues at NCEP), or something other than the CONDUIT data feed to distribute this data, or ??  It has come to the point where we can't rely on this data for plotting maps or doing analysis for classes, initializing local models, etc, because it has become so reliably incomplete, for the GFS runs in particular. 

Pete
https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+idd-agg.aos.wisc.edu <https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+idd-agg.aos.wisc.edu> 


<http://www.weather.com/tv/shows/wx-geeks/video/the-incredible-shrinking-cold-pool>-----
Pete Pokrandt - Systems Programmer
UW-Madison Dept of Atmospheric and Oceanic Sciences
608-262-3086  - address@hidden


------------------------------------------------------------------------
*From:* Tyle, Kevin R <address@hidden>
*Sent:* Friday, September 3, 2021 12:06 PM
*To:* Pete Pokrandt <address@hidden>; Jesse Marks - NOAA Affiliate <address@hidden>; address@hidden <address@hidden> *Cc:* _NWS NCEP NCO Dataflow <address@hidden>; Anne Myckow - NOAA Federal <address@hidden>; address@hidden <address@hidden> *Subject:* RE: [conduit] 20210830: Re: High CONDUIT latencies from vm-lnx-conduit2.ncep.noaa.gov 

Hi all,
After a few good days, we are once again not receiving all GFS forecast hours, starting with today’s 0000 UTC cycle. Pete, do you note the usual pattern of increasing latency from NCEP? 

Cheers,

Kevin

_________________________________________________

Kevin Tyle, M.S.; Manager of Departmental Computing

NSF XSEDE Campus Champion
Dept. of Atmospheric & Environmental Sciences
UAlbany ETEC Bldg – Harriman Campus
1220 Washington Avenue, Room 419
Albany, NY 12222
address@hidden | 518-442-4578 | @nywxguy | he/him/his

_________________________________________________
*From:* conduit <address@hidden> *On Behalf Of *Pete Pokrandt via conduit 
*Sent:* Tuesday, August 31, 2021 12:47 PM
*To:* Jesse Marks - NOAA Affiliate <address@hidden>; address@hidden
*Cc:* _NWS NCEP NCO Dataflow <address@hidden>; Anne Myckow - NOAA Federal <address@hidden>; address@hidden; address@hidden *Subject:* Re: [conduit] 20210830: Re: High CONDUIT latencies from vm-lnx-conduit2.ncep.noaa.gov
Thanks for the update, Jesse. I can confirm that we are seeing smaller lags originating from conduit2, and since yesterday's 18 UTC run, I don't think we have missed any data here at UW-Madison.
Kevin Tyle, how's your reception been at Albany since the 18 UTC run yesterday? 

Pete

-----
Pete Pokrandt - Systems Programmer
UW-Madison Dept of Atmospheric and Oceanic Sciences
608-262-3086  - address@hidden <mailto:address@hidden>

------------------------------------------------------------------------
*From:*Jesse Marks - NOAA Affiliate <address@hidden <mailto:address@hidden>> 
*Sent:* Tuesday, August 31, 2021 10:26 AM
*To:* address@hidden <mailto:address@hidden> <address@hidden <mailto:address@hidden>> *Cc:* Pete Pokrandt <address@hidden <mailto:address@hidden>>; Anne Myckow - NOAA Federal <address@hidden <mailto:address@hidden>>; address@hidden <mailto:address@hidden> <address@hidden <mailto:address@hidden>>; address@hidden <mailto:address@hidden> <address@hidden <mailto:address@hidden>>; _NWS NCEP NCO Dataflow <address@hidden <mailto:address@hidden>> *Subject:* Re: 20210830: Re: High CONDUIT latencies from vm-lnx-conduit2.ncep.noaa.gov
Thanks for the quick reply, Tom.  Looking through our conduit2 logs, we began seeing sends of product from our conduit2 to conduit1 machine after we restarted the LDM server on conduit2 yesterday.  It appears latencies improved fairly significantly at that time:
However we still do not see direct sends from conduit2 to external LDMs.  Our server team is currently looking into the TCP service issue that appears to be causing this problem. 

Thanks,

Jesse
On Mon, Aug 30, 2021 at 7:49 PM Tom Yoksas <address@hidden <mailto:address@hidden>> wrote: 

    Hi Jesse,

    On 8/30/21 5:16 PM, Jesse Marks - NOAA Affiliate wrote:
     > Quick question:  how are you computing these latencies?

    Latency in the LDM/IDD context is the time difference between when a
    product is first put into an LDM queue for redistribution and the time
    it is received by a downstream machine.  This measure of latency, of
    course, requires that the clocks on the originating and receiving
    machines be maintained accurately.

    re:
     > More
     > specifically, how do you determine which conduit machine the data is
     > coming from?

    The machine on which the product is inserted into the LDM queue is
    available in the LDM transaction.  We provide an website where users
    can create graphs of things like feed latencies:

    Unidata HomePage
    https://www.unidata.ucar.edu <https://www.unidata.ucar.edu>

        IDD Operational Status
    https://rtstats.unidata.ucar.edu/rtstats/
    <https://rtstats.unidata.ucar.edu/rtstats/>

          Real-time IDD Statistics -> Statistics by Host
    https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex
    <https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex>

    The variety of measures of feed quality for the Unidata machine that
    is REQUESTing the CNODUIT feed from the NCEP cluster can be found in:

    
https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex?conduit.unidata.ucar.edu
    
<https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex?conduit.unidata.ucar.edu>

    The latencies being reported by the Unidata machine that is being fed
    from the NCEP cluster is:

    CONDUIT latencies on conduit.conduit.unidata.ucar.edu
    <http://conduit.conduit.unidata.ucar.edu>:

    
https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+conduit.unidata.ucar.edu
    
<https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+conduit.unidata.ucar.edu>

    As you can see, the traces are color color coded, and the label at the
    top identifies the source machines for products.

    re:
     > The reason I ask is because I am not seeing any sends of
     > product from conduit2 in the last several days of logs both to
    our local
     > conduit1 machine and to any distant end users.

    Hmm...  we are.

    re:
     > Also, we have isolated what is likely the issue and will have our
    team
     > take a closer look in the morning.  I'm hopeful they'll be able to
     > resolve this soon.

    Excellent!  We are hopeful that the source of the high latencies will
    be identified and fixed.

    Cheers,

    Tom

     > On Mon, Aug 30, 2021 at 5:24 PM Anne Myckow - NOAA Federal
     > <address@hidden <mailto:address@hidden>
    <mailto:address@hidden <mailto:address@hidden>>> wrote:
     >
     >     Pete,
     >
     >     Random aside, can you please update your doco to say that
     >     Dataflow's email list is now address@hidden
    <mailto:address@hidden>
     >     <mailto:address@hidden <mailto:address@hidden>>
    ? I'm CC'ing it here. That other
     >     email address is going to get turned off within the next year.
     >
     >     Thanks,
     >     Anne
     >
     >     On Wed, Aug 18, 2021 at 4:02 PM Pete Pokrandt
    <address@hidden <mailto:address@hidden>
     >     <mailto:address@hidden <mailto:address@hidden>>> wrote:
     >
     >         Dear Anne, Dustin and all,
     >
     >         Recently we have noticed fairly high latencies on the CONDUIT
     >         ldm data feed originating from the machine
     > vm-lnx-conduit2.ncep.noaa.gov <http://vm-lnx-conduit2.ncep.noaa.gov>
     >         <http://vm-lnx-conduit2.ncep.noaa.gov
    <http://vm-lnx-conduit2.ncep.noaa.gov>>. The feed originating
     >         from vm-lnx-conduit1.ncep.noaa.gov
    <http://vm-lnx-conduit1.ncep.noaa.gov>
     >         <http://vm-lnx-conduit1.ncep.noaa.gov
    <http://vm-lnx-conduit1.ncep.noaa.gov>> does not have the high
     >         latencies. Unidata and other top level feeds are seeing
    similar
     >         high latencies from vm-lnx-conduit2.ncep.noaa.gov
    <http://vm-lnx-conduit2.ncep.noaa.gov>
     >         <http://vm-lnx-conduit2.ncep.noaa.gov
    <http://vm-lnx-conduit2.ncep.noaa.gov>>.
     >
     >         Here are some graphs showing the latencies that I'm seeing:
     >
     >          From
     >
    
https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+idd-agg.aos.wisc.edu
    
<https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+idd-agg.aos.wisc.edu>
>  <https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+idd-agg.aos.wisc.edu <https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/iddstats_nc?CONDUIT+idd-agg.aos.wisc.edu>> - 
     >         latencies for CONDUIT data arriving at our UW-Madison AOS
    ingest
     >         machine
     >
     >
     >
     >          From
     >
    
https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex?conduit.unidata.ucar.edu
    
<https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex?conduit.unidata.ucar.edu>
>  <https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex?conduit.unidata.ucar.edu <https://rtstats.unidata.ucar.edu/cgi-bin/rtstats/siteindex?conduit.unidata.ucar.edu>> (latencies 
     >         at Unidata)
     >
     >
     >
     >         At least here at UW-Madison, these latencies are causing
    us to
     >         lose some data during the large GFS/GEFS periods.
     >
     >         Any idea what might be causing this?
     >
     >         Pete
     >
     >
     >
     >
>  <http://www.weather.com/tv/shows/wx-geeks/video/the-incredible-shrinking-cold-pool <http://www.weather.com/tv/shows/wx-geeks/video/the-incredible-shrinking-cold-pool>>----- 
     >         Pete Pokrandt - Systems Programmer
     >         UW-Madison Dept of Atmospheric and Oceanic Sciences
     >         608-262-3086  - address@hidden
    <mailto:address@hidden> <mailto:address@hidden
    <mailto:address@hidden>>
     >
     >
     >
     >     --
     >     Anne Myckow
     >     Dataflow Team Lead
     >     NWS/NCEP/NCO
     >
     >
     >
     > --
     > Jesse Marks
     > Dataflow Analyst
     > NCEP Central Operations
     > 678-896-9420
-- +----------------------------------------------------------------------+ 
    * Tom Yoksas                                      UCAR Unidata Program *
    * (303) 497-8642 (last resort)                           P.O. Box 3000 *
* address@hidden <mailto:address@hidden>           Boulder, CO 80307 * 
    * Unidata WWW Service http://www.unidata.ucar.edu/
    <http://www.unidata.ucar.edu/> *
    +----------------------------------------------------------------------+



--

Jesse Marks

Dataflow Analyst

NCEP Central Operations

678-896-9420



--
+----------------------------------------------------------------------+
* Tom Yoksas                                      UCAR Unidata Program *
* (303) 497-8642 (last resort)                           P.O. Box 3000 *
* address@hidden                                    Boulder, CO 80307 *
* Unidata WWW Service                     http://www.unidata.ucar.edu/ *
+----------------------------------------------------------------------+