Le jeu., mai 25 2023 at 08:46:31 +0100, Andy Seaborne <[email protected]>
a écrit :
On 24/05/2023 10:22, Steven Blanchard wrote:
Hi Andy,
I tried it on a local disk and it had no impact on the average speed
for the Data stage.
SSD or rotating disk? (It shouldn't make an extreme difference or
xloader, because that's part of the point of the xloader.)
On a SSD. The average speed are the same on the SSD or on the Block
Storage.
I checked with iostat, there was indeed an increase in the speed of
reading the input files. This step writes very little data so there
was no difference in the writing speed.
I also did a test with only 1 of the uniprot files (291 million
tuples) and the average speed was about 160,000 tuples/s. This
value corresponds to speeds obtained on other insertions.
On the exact same hardware?
Yes same hardware, same folder, same time. Only the quantity of data is
differents.
Could this decrease of average speed be related to the amount of
total data?
Is it possible to run this Data step only file by file and all the
other steps with all files?
Not sure - there is a shared node table being built. The slowness is
presumably a consequence of the previous stages. The use of the some
URI needs to have the same internal NodeId everywhere - i.e. seeing
all the data.
During our tests, we resumed the insertion at the Data stage and we
noticed that the decrease in average speed is related to the previous
steps. If we give an argument the existing directory with the step
Nodes and Term having already instead the speed of insertion of the
data is 800 tuples/s. If we give an argument an empty directory, the
insertion speed of the Data step is 190,000 tuples/s.
The decrease in speed therefore seems to be related to the amount of
data and the results of the previous steps. When this step ingest data,
there is an optional step that uses previously created files that could
be very long because of the total amount of data?
What is the link between these 3 steps? What does each of these three
steps do for the data insertion?
I'm still not seeing why the data stage starts at a slow rate - I
will need to find time to explore the code.
(This is an argument for having NodeIds be hashes because that can be
computed without reference to the table unique ids and representation
storage. Downside - the NodeIds would be longer, 96 or 128 bits and
hashes have bad locality (i.e. none whatsoever)).
Andy
Thank you,
Steven
Le mar., mai 23 2023 at 11:30:36 +0100, Andy Seaborne
<[email protected] <mailto:[email protected]>> a écrit :
On 22/05/2023 16:38, Steven Blanchard wrote:
Le lun., mai 22 2023 at 16:18:21 +0100, Andy Seaborne
<[email protected] <mailto:[email protected]>
<<mailto:[email protected]>>> a écrit :
Hello Andy,
Hi Steven,
How are you runnign xloader? Default settings?
Yes, we use your default settings.
The command line used is the following line :
tdb2.xloader --loc /nfs/uniprot_tmp/tdb2/UniProt_04_2022/ --tmpdir
/nfs/uniprot_tmp/ --threads 30
/nfs/uniprot_tmp/Download/2022_04/uniprotkb_*.rdf
Just looking at that, the use of NFS may be related.
NFS is shared, remote filing system so it has comparative high
overheads on every operation to give the semantics of sharing
(visibility on write).
Could you try using local disk to see if that makes a difference?
Andy
What's the storage being used?
We use a Block Storage from a cloud providers with ssd on a mouted
nsf volume.
On 22/05/2023 10:49, Steven Blanchard wrote:
Hello,
I am currently trying to load a very large dataset ( 54 billion
triples) with the tdb2.xloader command.
The first two steps (Nodes and Terms) are completed with an
average load speed of ~ 120,000.
The third stage (Data) has an average load speed of only 800.
is thet "Avg" is 800 from teh start of the phase or "the average
drops to 800" during the phase?
The Avg is 800 from the start of the phase and he stay at 800.
This average load speed is incompatible with the amount of data
to be loaded.
Looking at the status of the job, it is possible that there is
an excessive demand on memory which slows down the
process extremely.
We saw with a top that java required many memories :
```
top
# PID USER PR NI VIRT RES SHR S
%CPU %MEM TIME+ COMMAND
# 867362 sblanch+ 20 0 289,0g 90,2g 88,4g S 3,3 72,1
1102:32 java
```
xloader does not have much requirement for java heap memory.
Ok, since our email we have try to increase the -xmx and we have
not an increase of the performance.
That space may be mapped files.
But with a free -g, we see that it actually uses very little
memory.
```
free -g
# total used free shared buff/cache available
# Mem: 125 3 0 0 121
120
```
Are there any possibilities to speed up this step? (Give a -xms
to java?)
Can this significant drop in loading speed for this step be due
to memory usage? Do you know of any other limiting
causes in this loading stage?
For previous insertions on smaller datasets, this Data step was
not limiting and the average speed was even slightly
higher than the Nodes and Terms steps.
How small is "smaller"?
For example, we have upload UniRef RDF Database (Same providers
like UniProt) with 12 Milliards of triples with an average
for Data task of 230 000 tuples/s
That sounds like what I see when loading.
For information, the machine used has 32 CPUs and 128 Giga of
Ram.
Thanks for your help,
Regards,
Steven
