<div dir="ltr">Hi All, <br>Adding my two cents, We have two kinds of storage First based on SSD 600TB and 2nd on spinning disks. 20PB.<br><br>for SSD i tried glusterfs ,beefs and compare them with lustre . Beefs failed because of ln (hard link issues) at that time (2019) . Glusterfs was very promising but somehow the lustre with zfs showed good performance. <br><br>on our 20PB we are running lustre based on zfs and happy with it. <br><br>I am trying to write a whitepaper with different parameters to tweak the lustre performance from disks, to zfs to lustre . This could help the HPC community, especially in life science. <br><br>/Zee<div>Section head of IT Infrastructure, </div><div>Centre for Genomic Medicine , KFSHRC , Riyadh<br><br></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Mon, Apr 5, 2021 at 6:54 PM Olaf Buitelaar <<a href="mailto:olaf.buitelaar@gmail.com">olaf.buitelaar@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr">Hi Erik,<div><br></div><div>Thanks for sharing your unique use-case and solution. It was very interesting to read your write-up.</div><div><br></div><div>I agree with your point; " * Growing volumes, replacing bricks, and replacing servers work.</div> However, the process is very involved and quirky for us. I have....." in your use-case 1 last point.<div><br></div><div>I do seem to suffer from similar issues where glusterd just doesn't want to start up correctly at first time, maybe also see: <a href="https://lists.gluster.org/pipermail/gluster-users/2021-February/039134.html" target="_blank">https://lists.gluster.org/pipermail/gluster-users/2021-February/039134.html</a> as one possible cause for not starting up correctly.</div><div>And secondly indeed for a "molten to the floor" server it would be great if gluster instead of the current replace-brick commands, would have something to replace a complete host, and would just recreate the all bricks (or better yet all missing bricks, in case some survived from another RAID/disk), which it would expect on that host, and would heal them all.</div><div>right now indeed this process is quite involved, and sometimes feels a bit like performing black magic.</div><div><br></div><div>Best Olaf</div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">Op vr 19 mrt. 2021 om 16:10 schreef Erik Jacobson <<a href="mailto:erik.jacobson@hpe.com" target="_blank">erik.jacobson@hpe.com</a>>:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">A while back I was asked to make a blog or something similar to discuss<br>
the use cases the team I work on (HPCM cluster management) at HPE.<br>
<br>
If you are not interested in reading about what I'm up to, just delete<br>
this and move on.<br>
<br>
I really don't have a public blogging mechanism so I'll just describe<br>
what we're up to here. Some of this was posted in some form in the past.<br>
Since this contains the raw materials, I could make a wiki-ized version<br>
if there were a public place to put it.<br>
<br>
<br>
<br>
We currently use gluster in two parts of cluster management.<br>
<br>
In fact, gluster in our management node infrastructure is helping us to<br>
provide scaling and consistency to some of the largest clusters in the<br>
world, clusters in the TOP100 list. While I can get in to trouble by<br>
sharing too much, I will just say that trends are continuing and the<br>
future may have some exciting announcements on where on TOP100 certain<br>
new giant systems may end up in the coming 1-2 years.<br>
<br>
At HPE, HPCM is the "traditional cluster manager." There is another team<br>
that develops a more cloud-like solution and I am not discussing that<br>
solution here.<br>
<br>
<br>
Use Case #1: Leader Nodes and Scale Out<br>
------------------------------------------------------------------------------<br>
- Why?<br>
* Scale out<br>
* Redundancy (combined with CTDB, any leader can fail)<br>
* Consistency (All servers and compute agree on what the content is)<br>
<br>
- Cluster manager has an admin or head node and zero or more leader nodes<br>
<br>
- Leader nodes are provisioned in groups of 3 to use distributed<br>
replica-3 volumes (although at least one customer has interest<br>
in replica-5)<br>
<br>
- We configure a few different volumes for different use cases<br>
<br>
- We use Gluster NFS still because, over a year ago, Ganesha was not<br>
working with our workload and we haven't had time to re-test and<br>
engage with the community. No blame - we would also owe making sure<br>
our settings are right.<br>
<br>
- We use CTDB for a measure of HA and IP alias management. We use this<br>
instead of pacemaker to reduce complexity.<br>
<br>
- The volume use cases are:<br>
* Image sharing for diskless compute nodes (sometimes 6,000 nodes)<br>
-> Normally squashFS image files for speed/efficiency exported NFS<br>
-> Expanded ("chrootable") traditional NFS trees for people who<br>
prefer that, but they don't scale as well and are slower to boot<br>
-> Squashfs images sit on a sharded volume while traditional gluster<br>
used for expanded tree.<br>
* TFTP/HTTP for network boot/PXE including miniroot<br>
-> Spread across leaders too due so one node is not saturated with<br>
PXE/DHCP requests<br>
-> Miniroot is a "fatter initrd" that has our CM toolchain<br>
* Logs/consoles<br>
-> For traditional logs and consoles (HCPM also uses<br>
elasticsearch/kafka/friends but we don't put that in gluster)<br>
-> Separate volume to have more non-cached friendly settings<br>
* 4 total volumes used (one sharded, one heavily optimized for<br>
caching, one for ctdb lock, and one traditional for logging/etc)<br>
<br>
- Leader Setup<br>
* Admin node installs the leaders like any other compute nodes<br>
* A setup tool operates that configures gluster volumes and CTDB<br>
* When ready, an admin/head node can be engaged with the leaders<br>
* At that point, certain paths on the admin become gluster fuse mounts<br>
and bind mounts to gluster fuse mounts.<br>
<br>
- How images are deployed (squashfs mode)<br>
* User creates an image using image creation tools that make a<br>
chrootable tree style image on the admin/head node<br>
* mksquashfs will generate a squashfs image file on to a shared<br>
storage gluster mount<br>
* Nodes will mount the filesystem with the squashfs images and then<br>
loop mount the squashfs as part of the boot process.<br>
<br>
- How are compute nodes tied to leaders<br>
* We simply have a variable in our database where human or automated<br>
discovery tools can assign a given node to a given IP alias. This<br>
works better for us than trying to play routing tricks or load<br>
balance tricks<br>
* When leaders PXE, the DHCP response includes next-server and the<br>
compute node uses the leader IP alias for the tftp/http for<br>
getting the boot loader DHCP config files are on shared storage<br>
to facilitate future scaling of DHCP services.<br>
* ipxe or grub2 network config files then fetch the kernel, initrd<br>
* initrd has a small update to load a miniroot (install environment)<br>
which has more tooling<br>
* Node is installed (for nodes with root disks) or does a network boot<br>
cycle.<br>
<br>
- Gluster sizing<br>
* We typically state compute nodes per leader but this is not for<br>
gluster per-se. Squashfs image objects are very efficient and<br>
probably would be fine for 2k nodes per leader. Leader nodes provide<br>
other services including console logs, system logs, and monitoring<br>
services.<br>
* Our biggest deployment at a customer site right now has 24 leader<br>
nodes. Bigger systems are coming.<br>
<br>
- Startup scripts - Getting all the gluster mounts and many bind mounts<br>
used in the solution, as well as ensuring gluster mounts and ctdb lock<br>
is available before ctdb start was too painful for my brian. So we<br>
have systemd startup scripts that sanity test and start things<br>
gracefully.<br>
<br>
- Future: The team is starting to test what a 96-leader (96 gluster<br>
servers) might look like for future exascale systems.<br>
<br>
- Future: Some customers have interest in replica-5 instead of<br>
replica-3. We want to test performance implications.<br>
<br>
- Future: Move to Ganesha, work with community if needed<br>
<br>
- Future: Work with Gluster community to make gluster fuse mounts<br>
efficient instead of NFS (may be easier with image objects than it was<br>
the old way with fully expanded trees for images!)<br>
<br>
- Challenges:<br>
* Every month someone tells me gluster is going to die because of Red<br>
Hat vs IBM and I have to justify things. It is getting harder.<br>
* Giant squashfs images fail - mksquashfs reports error - at around<br>
90GB on sles15sp2 and sles15sp3. rhel8 does not suffer. Don't have<br>
the bandwidth to dig in right now but one upset customer. Work<br>
arounds provided to move development environment for that customer<br>
out of the operating system image.<br>
* Since we have our own build and special use cases, we're on our own<br>
for support (by "on our own" I mean no paid support, community help<br>
only). Our complex situations can produce some cases that you guys<br>
don't see and debugging them can take a month or more with the<br>
volunteer nature of the community. Paying for support is harder,<br>
even if it were viable politically, since we support 6 distros and 3<br>
distro providers. Of course, paying for support is never the<br>
preference of management. It might be an interesting thing to<br>
investigate.<br>
* Any gluster update is terror. We don't update much because finding a<br>
gluster version that is stable for all our use cases PLUS being able to<br>
test at scale which means thousands of nodes is hard. We did some<br>
internal improvements here where we emulate a 2500-node-cluster<br>
using virtual machines on a much smaller cluster. However, it isn't<br>
ideal. So we start lagging the community over time until some<br>
problem forces us to refresh. Then we tip-toe in to the update. We<br>
most recently updated to gluster79 and it solved several problems<br>
related to use case #2 below.<br>
* Due to lack of HW, testing the SU Leader solution is hard because of<br>
the number of internal clusters. However, I recently moved my<br>
primary development to my beefy desktop where I have a full cluster<br>
stack including 3 leader nodes with gluster running in virtual<br>
machines. So we have eliminated an excuse preventing internal people<br>
from playing with the solution stack.<br>
* Growing volumes, replacing bricks, and replacing servers work.<br>
However, the process is very involved and quirky for us. I have<br>
complicated code that has to do more than I'd really like to do to<br>
simply wheel in a complete leader replacement for a failed one. Even<br>
with our tools, we often send up with some glusterd's not starting<br>
right and have to restart a few times to get a node or brick<br>
replacement going. I wish the process were just as simple as running<br>
a single command or set of commands and having it do the right<br>
thing.<br>
-> Finding two leaders to get a complete set of peer files<br>
-> Wedging in the replacement node with the UUID the node in that<br>
position had before<br>
-> Don't accidentally give a gluster peer it's own peer file<br>
-> Then go through an involved replace brick procedure<br>
-> "replace" vs " reset"<br>
-> A complicated dance with XATTRs that I still don't understand<br>
-> Ensuring indices and .glusterfs pre-exist with right<br>
permissions<br>
My hope is I just misunderstand and will bring this up in a future<br>
discussion.<br>
<br>
<br>
<br>
<br>
Use Case #2: Admin/Head Node High Availability<br>
------------------------------------------------------------------------------<br>
- We used to have an admin node HA solution that contained two servers<br>
and an external storage device. A VM was used for the "real admin<br>
node" provided by the two servers.<br>
<br>
- This solution was expensive due to the external storage<br>
<br>
- This solution was less optimal due to not having true quorum<br>
<br>
- Building on our gluster experience, we removed the external storage<br>
and added a 3rd server.<br>
(Due to some previous experience with DRBD we elected to stay with<br>
gluster here)<br>
<br>
- We build a gluster volume shared across the 3 servers, sharded, which<br>
primarily holds a virtual machine image file used by the admin node VM<br>
<br>
- The physical nodes use bonding for network redundancy and bridges to<br>
feed them in to the VM.<br>
<br>
- We use pacemaker in this case to manage the setup<br>
<br>
- It's pretty simple - pacemaker rules manage a VirtualDomain instance<br>
and a simple ssh monitor makes sure we can get in to it.<br>
<br>
- We typically have a 3-12T single image sitting on gluster sharded<br>
shared storage used by the virtual machine, which forms the true admin<br>
node.<br>
<br>
- We set this up with manual instructions but tooling is coming soon to<br>
aid in automated setup of this solution<br>
<br>
- This solution is in use actively in at least 4 very large<br>
supercomputers.<br>
<br>
- I am impressed by the speed of the solution on the sharded volume. The<br>
VM image creation speed using libvirt to talk to the image file hosted<br>
on a sharded gluster volume works slick.<br>
(We use the fuse mount because we don't want to build our own<br>
qemu/libvirt, which would be needed at least for SLES and maybe<br>
RHEL too since we have our own gluster build)<br>
<br>
- Challenges:<br>
* Not being able to boot all of a sudden was a terror for us (where<br>
the VM would only see the disk size as the size of a shard at random<br>
times).<br>
-> Thankfully community helped guide us to gluster79 and that resolved it<br>
<br>
* People keep asking to make a 2-node version but I push back.<br>
Especially with gluster but honestly with other solutions too, don't<br>
cheap out with arbitration is what I try to tell people.<br>
<br>
<br>
Network Utilization<br>
------------------------------------------------------------------------------<br>
- We tyhpically have 2x 10G bonds on leaders<br>
- For simplicity, we co-mingle gluster and compute node traffic together<br>
- It is very rare that we approach 20G full utilization even in very<br>
intensive operations<br>
- Newer solutions may increase the speed of the bonds, but it isn't due<br>
to a current pressing need.<br>
- Locality Challenge:<br>
* For future Exascale systems, we may need to become concerned about<br>
locality<br>
* Certain compute nodes are far closer to some gluster servers than<br>
others<br>
* And gluster servers themselves need to talk among themselves but<br>
could be stretched across the topology<br>
* We have tools to monitor switch link utilization and so far have<br>
not hit a scary zone<br>
* Somewhat complicated by fault tolerance. It would be sad to design<br>
the leaders such that a PDU goes bad so you lose qourum because<br>
3 leaders in the same replica-3 were on the same PDU<br>
* But spreading them out may have locality implications<br>
* This is a future area of study for us. We have various ideas in<br>
mind if a problem strikes.<br>
________<br>
<br>
<br>
<br>
Community Meeting Calendar:<br>
<br>
Schedule -<br>
Every 2nd and 4th Tuesday at 14:30 IST / 09:00 UTC<br>
Bridge: <a href="https://meet.google.com/cpu-eiue-hvk" rel="noreferrer" target="_blank">https://meet.google.com/cpu-eiue-hvk</a><br>
Gluster-users mailing list<br>
<a href="mailto:Gluster-users@gluster.org" target="_blank">Gluster-users@gluster.org</a><br>
<a href="https://lists.gluster.org/mailman/listinfo/gluster-users" rel="noreferrer" target="_blank">https://lists.gluster.org/mailman/listinfo/gluster-users</a><br>
</blockquote></div>
________<br>
<br>
<br>
<br>
Community Meeting Calendar:<br>
<br>
Schedule -<br>
Every 2nd and 4th Tuesday at 14:30 IST / 09:00 UTC<br>
Bridge: <a href="https://meet.google.com/cpu-eiue-hvk" rel="noreferrer" target="_blank">https://meet.google.com/cpu-eiue-hvk</a><br>
Gluster-users mailing list<br>
<a href="mailto:Gluster-users@gluster.org" target="_blank">Gluster-users@gluster.org</a><br>
<a href="https://lists.gluster.org/mailman/listinfo/gluster-users" rel="noreferrer" target="_blank">https://lists.gluster.org/mailman/listinfo/gluster-users</a><br>
</blockquote></div>