<div dir="ltr">Hi Valdis,<div><br></div><div>I tired with the grep replaced with 'ls -ls' and 'md5sum', I don't think this is a data integrity issue, thankfully:</div><div><br></div><div><div>$ ./pipetestls.sh </div><div>256 -rw-r--r-- 1 39073 3001 530721 Feb 14 07:16 /srv/gsfs0/projects/pipetest.tmp.txt</div><div>0 -rw-r--r-- 1 39073 3953 530721 Feb 14 07:16 /home/griznog/pipetest.tmp.txt</div><div><br></div><div>$ ./pipetestmd5.sh </div><div>15cb81a85c9e450bdac8230309453a0a /srv/gsfs0/projects/pipetest.tmp.txt</div><div>15cb81a85c9e450bdac8230309453a0a /home/griznog/pipetest.tmp.txt</div></div><div><br></div><div>And replacing grep with 'file' even properly sees the files as ASCII:</div><div><div>$ ./pipetestfile.sh </div><div>/srv/gsfs0/projects/pipetest.tmp.txt: ASCII text, with very long lines</div><div>/home/griznog/pipetest.tmp.txt: ASCII text, with very long lines</div></div><div><br></div><div>I'll poke a little harder at grep next and see what the difference in strace of each reveals.</div><div><br></div><div>Thanks,</div><div><br></div><div>jbh</div><div><br></div><div><br></div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Feb 14, 2018 at 7:08 AM, <span dir="ltr"><<a href="mailto:valdis.kletnieks@vt.edu" target="_blank">valdis.kletnieks@vt.edu</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">On Wed, 14 Feb 2018 06:20:32 -0800, John Hanks said:<br>
<br>
> # ls -aln /srv/gsfs0/projects/pipetest.<wbr>tmp.txt $HOME/pipetest.tmp.txt<br>
> -rw-r--r-- 1 39073 3953 530721 Feb 14 06:10 /home/griznog/pipetest.tmp.txt<br>
> -rw-r--r-- 1 39073 3001 530721 Feb 14 06:10<br>
> /srv/gsfs0/projects/pipetest.<wbr>tmp.txt<br>
><br>
> We can "fix" the user case that exposed this by not using a temp file or<br>
> inserting a sleep, but I'd still like to know why GPFS is behaving this way<br>
> and make it stop.<br>
<br>
</span>May be related to replication, or other behind-the-scenes behavior.<br>
<br>
Consider this example - 4.2.3.6, data and metadata replication both<br>
set to 2, 2 sites 95 cable miles apart, each is 3 Dell servers with a full<br>
fiberchannel mesh to 3 Dell MD34something arrays.<br>
<br>
% dd if=/dev/zero bs=1k count=4096 of=sync.test; ls -ls sync.test; sleep 5; ls -ls sync.test; sleep 5; ls -ls sync.test<br>
4096+0 records in<br>
4096+0 records out<br>
4194304 bytes (4.2 MB) copied, 0.0342852 s, 122 MB/s<br>
2048 -rw-r--r-- 1 root root 4194304 Feb 14 09:35 sync.test<br>
8192 -rw-r--r-- 1 root root 4194304 Feb 14 09:35 sync.test<br>
8192 -rw-r--r-- 1 root root 4194304 Feb 14 09:35 sync.test<br>
<br>
Notice that the first /bin/ls shouldn't be starting until after the dd has<br>
completed - at which point it's only allocated half the blocks needed to hold<br>
the 4M of data at one site. 5 seconds later, it's allocated the blocks at both<br>
sites and thus shows the full 8M needed for 2 copies.<br>
<br>
I've also seen (but haven't replicated it as I write this) a small file (4-8K<br>
or so) showing first one full-sized block, then a second full-sized block, and<br>
then dropping back to what's needed for 2 1/32nd fragments. That had me<br>
scratching my head<br>
<br>
Having said that, that's all metadata fun and games, while your case<br>
appears to have some problems with data integrity (which is a whole lot<br>
scarier). It would be *really* nice if we understood the problem here.<br>
<br>
The scariest part is:<br>
<span class=""><br>
> The first grep | wc -l returns 1, because grep outputs "Binary file /path/to/<br>
> gpfs/mount/test matches"<br>
<br>
</span>which seems to be implying that we're failing on semantic consistency.<br>
Basically, your 'cat' command is completing and closing the file, but then a<br>
temporally later open of the same find is reading something other that only the<br>
just-written data. My first guess is that it's a race condition similar to the<br>
following: The cat command is causing a write on one NSD server, and the first<br>
grep results in a read from a *different* NSD server, returning the data that<br>
*used* to be in the block because the read actually happens before the first<br>
NSD server actually completes the write.<br>
<br>
It may be interesting to replace the grep's with pairs of 'ls -ls / dd' commands to grab the<br>
raw data and its size, and check the following:<br>
<br>
1) does the size (both blocks allocated and logical length) reported by<br>
ls match the amount of data actually read by the dd?<br>
<br>
2) Is the file length as actually read equal to the written length, or does it<br>
overshoot and read all the way to the next block boundary?<br>
<br>
3) If the length is correct, what's wrong with the data that's telling grep that<br>
it's a binary file? ( od -cx is your friend here).<br>
<br>
4) If it overshoots, is the remainder all-zeros (good) or does it return semi-random<br>
"what used to be there" data (bad, due to data exposure issues)?<br>
<br>
(It's certainly not the most perplexing data consistency issue I've hit in 4 decades - the<br>
winner *has* to be a intermittent data read corruption on a GPFS 3.5 cluster that<br>
had us, IBM, SGI, DDN, and at least one vendor of networking gear all chasing our<br>
tails for 18 months before we finally tracked it down. :)<br>
<br>______________________________<wbr>_________________<br>
gpfsug-discuss mailing list<br>
gpfsug-discuss at <a href="http://spectrumscale.org" rel="noreferrer" target="_blank">spectrumscale.org</a><br>
<a href="http://gpfsug.org/mailman/listinfo/gpfsug-discuss" rel="noreferrer" target="_blank">http://gpfsug.org/mailman/<wbr>listinfo/gpfsug-discuss</a><br>
<br></blockquote></div><br></div>