Sorry, I forgot to introduce myself.
My name is Txema Heredia. I am a systems administrator at the
Evolutionary Biology Institute in Barcelona, Spain. We are a public
research institution focused in biological and genetics research. We
have a small cluster (300-cores) and we use GPFS to feed it with a 150TB
filesystem that is currently being upgraded with ~450TB more.
We have been working with GPFS for less than a year. Our initial
installation was made by on-site IBM technicians. But we are upgrading
the system on our own, and now I am begining to understand the guts of
GPFS and all its nuances.
I looking forward to learn a lot from this discussion list.
Cheers,
Txema
El 14/06/13 19:57, Txema Heredia Genestar escribió:
Hi all,
We are building a new GPFS cluster and I have a few doubts that I hope
you can solve about the NSD threads.
Our old cluster is composed of 2 building blocks. Each BB is composed
by 2 servers (12-core with 48GB RAM) connected by SAS to a dual
controller DS3512 disk cabinet, with 36x 7.2krpm 3TB SATA disks. Each
controller has 2Gb of cache.
Our "big" filesystem (130 TB) is formed by 6 NSDs, each one being a
8+1 RAID5 LUN coming from a cabinet. Data and metadata mixed. We have
6 luns and 4 controllers and NSD servers. Thus, some serve 2 "disks"
and some just 1.
As for GPFS, we are using the default GPFS 3.4 thread parameters:
nsdMaxWorkerThreads = 64
nsdMinWorkerThreads = 16
nsdThreadsPerDisk = 3
#NSD per server = 1 or 2
In this an IBM presentation (
http://www-05.ibm.com/de/events/gpfs-workshop/pdf/pr-11-GPFS_R35_nsdMultipleQ_and_other_enhancmentsv4-OW.pdf
slide 4), they show that the formula to obtain the number of
concurrently active nsd threads is:
MAX ( MIN ( nsdThreadsPerDisk * #NSDperServer , nsdMaxWorkerThreads
), nsdMinWorkerThreads )
In our case, we have only 6 NSD, and a server is responsible only of
up-to 2 of them. We are left with MIN ( 6 , 16 ), and thus, we end up
having between 8 and 16 threads per disk, when we should have just 3.
This is a photo obtained right now in one of our servers with "mmfsadm
dump nsd":
Worker threads: running 16, started 16, desired 16, active 16, highest 16
Requests: pending 333, highest pending 615, total processed 839099802
Buffer use: current 16777216, highest 16777216
Server state: suspendCount 0, killRequested 0, activeLocalIO 0
reOpenRequested 0, reOpenInProgress 0, nsdJoinComplete 1,
osdRequests 0x0
[...]
Disk name NsdId Status Hold I/O rcktry wckerr Addr
---------- ----------------- -------- ---- --- ------ ------ ----
home11 0A3C3D02:4FC87656 active 0 0 0 0
0x7F4E501565C0
scratch11 0A3C3D01:4FBE76D7 active 15 15 0 0
0x7F4E50156640
scratch12 0A3C3D02:4FBE76D8 active 0 0 0 0
0x7F4E501566C0
scratch13 0A3C3D01:4FBE76D8 active 1 1 0 0
0x7F4E50156740
On the other hand, when we run the performance monitor on each disk
controller, we obtain the following numbers per LUN in a state of 0%
cache hit:
mean IO/s = 180
Read % = 97.5%
throughput = 105 MB/s
All LUNs show similar results. The combined read throughput is ~630
MB/s. This is is the "live" cluster with ~300 jobs running, not a
single process reading a big file.
Are all these numbers ok? Is that disk performance fine?
What should we do with the thread parameters? Are the 16 simultaneous
threads disrupting our disk? Should we lower the nsdMinWorkerThreads?
Are they not? Should we rise the nsdThreadsPerDisk to 16 or more, as
the disks have shown they can handle them?
In our new cluster installation, we will have 4 nsd servers, each one
being responsible of 4-5 NSDs, using 4 disk cabinets similar to the
ones we have now. We will also move to GPFS 3.5, where
nsdMaxWorkerThreads has been rised to 512 as default and with the
small/large queues thing.
How should we adapt to it? Is the nsdThreadsPerDisk=3 an ancient
default value and we should move on?
Thanks in advance,
Txema
_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at gpfsug.org
http://gpfsug.org/mailman/listinfo/gpfsug-discuss
