Hi Eugen,Thank you very much for these detailed tests that match what I observed and reported earlier. I'm happy to see that we have the same understanding of how it should work (based on the documentation). Is there any other way that this list to enter in contact with the plugin developers as it seems they are not following this (very high volume) list... Or may somebody pass the email thread to one of them?
Help would be really appreciated. Cheers, Michel Le 19/06/2023 à 14:09, Eugen Block a écrit :
Hi, I have a real hardware cluster for testing available now. I'm not sure whether I'm completely misunderstanding how it's supposed to work or if it's a bug in the LRC plugin. This cluster has 18 HDD nodes available across 3 rooms (or DCs), I intend to use 15 nodes to be able to recover if one node fails. Given that I need one additional locality chunk per DC I need a profile with k + m = 12. So I chose k=9, m=3, l=4 which creates 15 chunks in total across those 3 DCs, one chunk per host, I checked the chunk placement and it is correct. This is the profile I created:ceph osd erasure-code-profile set lrc1 plugin=lrc k=9 m=3 l=4 crush-failure-domain=host crush-locality=room crush-device-class=hddI created a pool with only one PG to make the output more readable.This profile should allow the cluster to sustain the loss of three chunks, the results are interesting. This is what I tested:1. I stopped all OSDs on one host and the PG was still active with one missing chunk, everything's good. 2. Stopping a second host in the same DC resulted in the PG being marked as "down". That was unexpected since with m=3 I expected the PG to still be active but degraded. Before test #3 I started all OSDs to have the PG active+clean again. 3. I stopped one host per DC, so in total 3 chunks were missing and the PG was still active.Apparently, this profile is able to sustain the loss of m chunks, but not an entire DC. I get the impression (and I also discussed this with a colleague) that LRC with this implementation is either designed to loose only single OSDs which can be recovered quicker with fewer surviving OSDs and saving bandwidth. Or this is a bug because according to the low-level description [1] the algorithm works its way up in the reverse order within the configured layers, like in this example (not displaying my k, m, l requirements, just for reference):chunk nr 01234567 step 1 _cDD_cDD step 2 cDDD____ step 3 ____cDDDSo if a whole DC fails and the chunks from step 3 can not be recovered, and maybe step 2 also fails, but eventually step 1 contains the actual k and m chunks which should sustain the loss of an entire DC. My impression is that the algorithm somehow doesn't arrive at step 1 and therefore the PG stays down although there are enough surviving chunks. I'm not sure if my observations and conclusion are correct, I'd love to have a comment from the developers on this topic. But in this state I would not recommend to use the LRC plugin when the resiliency requirements are to sustain the loss of an entire DC.Thanks, Eugen[1] https://docs.ceph.com/en/latest/rados/operations/erasure-code-lrc/#low-level-plugin-configurationZitat von Michel Jouvin <michel.jouvin@xxxxxxxxxxxxxxx>:Hi,I realize that the crushmap I attached to one of my email, probably required to understand the discussion here, has been stripped down by mailman. To avoid poluting the thread with a long output, I put it on at https://box.in2p3.fr/index.php/s/J4fcm7orfNE87CX. Download it if you are interested.Best regards, Michel Le 21/05/2023 à 16:07, Michel Jouvin a écrit :Hi Eugen,My LRC pool is also somewhat experimental so nothing really urgent. If you manage to do some tests that help me to understand the problem I remain interested. I propose to keep this thread for that.Zitat, I shared my crush map in the email you answered if the attachment was not suppressed by mailman.Cheers, Michel Sent from my mobile Le 18 mai 2023 11:19:35 Eugen Block <eblock@xxxxxx> a écrit :Hi, I don’t have a good explanation for this yet, but I’ll soon get the opportunity to play around with a decommissioned cluster. I’ll try to get a better understanding of the LRC plugin, but it might take some time, especially since my vacation is coming up. :-) I have some thoughts about the down PGs with failure domain OSD, but I don’t have anything to confirm it yet. Zitat von Curt <lightspd@xxxxxxxxx>:Hi,I've been following this thread with interest as it seems like a unique use case to expand my knowledge. I don't use LRC or anything outside basicerasure coding.What is your current crush steps rule? I know you made changes since your first post and had some thoughts I wanted to share, but wanted to see your rule first so I could try to visualize the distribution better. The only way I can currently visualize it working is with more servers, I'm thinking 6 or 9 per data center min, but that could be my lack of knowledge on someof the step rules. Thanks Curt On Tue, May 16, 2023 at 11:09 AM Michel Jouvin < michel.jouvin@xxxxxxxxxxxxxxx> wrote:Hi Eugen,Yes, sure, no problem to share it. I attach it to this email (as it mayclutter the discussion if inline). If somebody on the list has some clue on the LRC plugin, I'm still interested by understand what I'm doing wrong! Cheers, Michel Le 04/05/2023 à 15:07, Eugen Block a écrit :[NONE,236,28,228,218,NONE,215,117,203,213,204,115,136,181,171,162,137,128]Hi, I don't think you've shared your osd tree yet, could you do that?Apparently nobody else but us reads this thread or nobody reading thisuses the LRC plugin. ;-) Thanks, Eugen Zitat von Michel Jouvin <michel.jouvin@xxxxxxxxxxxxxxx>:Hi,I had to restart one of my OSD server today and the problem showed upagain. This time I managed to capture "ceph health detail" output showing the problem with the 2 PGs:[WRN] PG_AVAILABILITY: Reduced data availability: 2 pgs inactive, 2pgs down pg 56.1 is down, acting[208,65,73,206,197,193,144,155,178,182,183,133,17,NONE,36,NONE,230,NONE]pg 56.12 is down, actingI still doesn't understand why, if I am supposed to survive to adatacenter failure, I cannot survive to 3 OSDs down on the same host, hosting shards for the PG. In the second case it is only 2 OSDs downbut I'm surprised they don't seem in the same "group" of OSD (I'dexpected all the the OSDs of one datacenter to be in the same groupeof 5 if the order given really reflects the allocation done... Still interested by some explanation on what I'm doing wrong! Best regards, Michel Le 03/05/2023 à 10:21, Eugen Block a écrit :I think I got it wrong with the locality setting, I'm still limited by the number of hosts I have available in my test cluster, but asfar as I got with failure-domain=osd I believe k=6, m=3, l=3 with locality=datacenter could fit your requirement, at least with regards to the recovery bandwidth usage between DCs, but theresiliency would not match your requirement (one DC failure). That profile creates 3 groups of 4 chunks (3 data/coding chunks and one parity chunk) across three DCs, in total 12 chunks. The min_size=7 would not allow an entire DC to go down, I'm afraid, you'd have toreduce it to 6 to allow reads/writes in a disaster scenario. I'mstill not sure if I got it right this time, but maybe you're better off without the LRC plugin with the limited number of hosts. Insteadyou could use the jerasure plugin with a profile like k=4 m=5 allowing an entire DC to fail without losing data access (we have one customer using that). Zitat von Eugen Block <eblock@xxxxxx>:Hi,disclaimer: I haven't used LRC in a real setup yet, so there mightbe some misunderstandings on my side. But I tried to play aroundwith one of my test clusters (Nautilus). Because I'm limited in thenumber of hosts (6 across 3 virtual DCs) I tried two different profiles with lower numbers to get a feeling for how that works. # first attempt ceph:~ # ceph osd erasure-code-profile set LRCprofile plugin=lrc k=4 m=2 l=3 crush-failure-domain=hostFor every third OSD one parity chunk is added, so 2 more chunks to store ==> 8 chunks in total. Since my failure-domain is host and Ionly have 6 I get incomplete PGs. # second attempt ceph:~ # ceph osd erasure-code-profile set LRCprofile plugin=lrc k=2 m=2 l=2 crush-failure-domain=hostThis gives me 6 chunks in total to store across 6 hosts which works:ceph:~ # ceph pg ls-by-pool lrcpool PG OBJECTS DEGRADED MISPLACED UNFOUND BYTES OMAP_BYTES* OMAP_KEYS* LOG STATE SINCE VERSION REPORTED UP ACTING SCRUB_STAMP DEEP_SCRUB_STAMP 50.0 1 0 0 0 619 0 0 1 active+clean 72s 18410'1 18415:54 [27,13,0,2,25,7]p27 [27,13,0,2,25,7]p27 2023-05-02 14:53:54.322135 2023-05-02 14:53:54.322135 50.1 0 0 0 0 0 0 0 0 active+clean 6m 0'0 18414:26 [27,33,22,6,13,34]p27 [27,33,22,6,13,34]p27 2023-05-02 14:53:54.322135 2023-05-02 14:53:54.322135 50.2 0 0 0 0 0 0 0 0 active+clean 6m 0'0 18413:25 [1,28,14,4,31,21]p1 [1,28,14,4,31,21]p1 2023-05-02 14:53:54.322135 2023-05-02 14:53:54.322135 50.3 0 0 0 0 0 0 0 0 active+clean 6m 0'0 18413:24 [8,16,26,33,7,25]p8 [8,16,26,33,7,25]p8 2023-05-02 14:53:54.322135 2023-05-02 14:53:54.322135 After stopping all OSDs on one host I was still able to read andwrite into the pool, but after stopping a second host one PG fromthat pool went "down". That I don't fully understand yet, but I just started to look into it.With your setup (12 hosts) I would recommend to not utilize all ofthem so you have capacity to recover, let's say one "spare" hostper DC, leaving 9 hosts in total. A profile with k=3 m=3 l=2 couldmake sense here, resulting in 9 total chunks (one more parity chunks for every other OSD), min_size 4. But as I wrote, it probably doesn't have the resiliency for a DC failure, so that needs some further investigation. Regards, Eugen Zitat von Michel Jouvin <michel.jouvin@xxxxxxxxxxxxxxx>:Hi, No... our current setup is 3 datacenters with the sameconfiguration, i.e. 1 mon/mgr + 4 OSD servers with 16 OSDs each. Thus the total of 12 OSDs servers. As with LRC plugin, k+m must bea multiple of l, I found that k=9/m=66/l=5 withcrush-locality=datacenter was achieving my goal of being resilientto a datacenter failure. Because I had this, I considered thatlowering the crush failure domain to osd was not a major issue in my case (as it would not be worst than a datacenter failure if all the shards are on the same server in a datacenter) and was workingaround the lack of hosts for k=9/m=6 (15 OSDs). May be it helps, if I give the erasure code profile used: crush-device-class=hdd crush-failure-domain=osd crush-locality=datacenter crush-root=default k=9 l=5 m=6 plugin=lrcThe previously mentioned strange number for min_size for the poolcreated with this profile has vanished after Quincy upgrade as this parameter is no longer in the CRUH map rule! and the `ceph osd pool get` command reports the expected number (10): ---------ceph osd pool get fink-z1.rgw.buckets.data min_sizemin_size: 10 -------- Cheers, Michel Le 29/04/2023 à 20:36, Curt a écrit :Hello,What is your current setup, 1 server pet data center with 12 osdeach? What is your current crush rule and LRC crush rule? On Fri, Apr 28, 2023, 12:29 Michel Jouvin <michel.jouvin@xxxxxxxxxxxxxxx> wrote: Hi, I think I found a possible cause of my PG down but still understand why.As explained in a previous mail, I setup a 15-chunk/OSD EC pool(k=9, m=6) but I have only 12 OSD servers in the cluster. To workaround theproblem I defined the failure domain as 'osd' with the reasoningthat asI was using the LRC plugin, I had the warranty that I could loosea sitewithout impact, thus the possibility to loose 1 OSD server. Am Iwrong? Best regards, Michel Le 24/04/2023 à 13:24, Michel Jouvin a écrit : > Hi, > > I'm still interesting by getting feedback from those using the LRC > plugin about the right way to configure it... Last week I upgraded> from Pacific to Quincy (17.2.6) with cephadm which is doing the > upgrade host by host, checking if an OSD is ok to stop beforeactually> upgrading it. I had the surprise to see 1 or 2 PGs down at somepoints > in the upgrade (happened not for all OSDs but for every > site/datacenter). Looking at the details with "ceph health detail", I> saw that for these PGs there was 3 OSDs down but I was expectingthe> pool to be resilient to 6 OSDs down (5 for R/W access) so I'm > wondering if there is something wrong in our pool configuration(k=9, > m=6, l=5). > > Cheers, > > Michel > > Le 06/04/2023 à 08:51, Michel Jouvin a écrit : >> Hi, >> >> Is somebody using LRC plugin ? >>>> I came to the conclusion that LRC k=9, m=3, l=4 is not thesame as>> jerasure k=9, m=6 in terms of protection against failures andthat I >> should use k=9, m=6, l=5 to get a level of resilience >= jerasure>> k=9, m=6. The example in the documentation (k=4, m=2, l=3)suggests >> that this LRC configuration gives something better than jerasure k=4,>> m=2 as it is resilient to 3 drive failures (but not 4 if Iunderstood>> properly). So how many drives can fail in the k=9, m=6, l=5>> configuration first without loosing RW access and second without >> loosing data? >>>> Another thing that I don't quite understand is that a poolcreated >> with this configuration (and failure domain=osd, locality=datacenter)>> has a min_size=3 (max_size=18 as expected). It seems wrong tome, I'd>> expected something ~10 (depending on answer to the previousquestion)... >> >> Thanks in advance if somebody could provide some sort of >> authoritative answer on these 2 questions. Best regards, >> >> Michel >> >> Le 04/04/2023 à 15:53, Michel Jouvin a écrit :>>> Answering to myself, I found the reason for 2147483647: it's >>> documented as a failure to find enough OSD (missing OSDs). Andit is>>> normal as I selected different hosts for the 15 OSDs but Ihave only >>> 12 hosts! >>>>>> I'm still interested by an "expert" to confirm that LRC k=9,m=3,>>> l=4 configuration is equivalent, in terms of redundancy, to a>>> jerasure configuration with k=9, m=6. >>> >>> Michel >>> >>> Le 04/04/2023 à 15:26, Michel Jouvin a écrit : >>>> Hi, >>>> >>>> As discussed in another thread (Crushmap rule for multi-datacenter>>>> erasure coding), I'm trying to create an EC pool spanning 3>>>> datacenters (datacenters are present in the crushmap), with the>>>> objective to be resilient to 1 DC down, at least keeping the >>>> readonly access to the pool and if possible the read-writeaccess,>>>> and have a storage efficiency better than 3 replica (letsay a >>>> storage overhead <= 2). >>>> >>>> In the discussion, somebody mentioned LRC plugin as a possible>>>> jerasure alternative to implement this without tweaking the>>>> crushmap rule to implement the 2-step OSD allocation. I looked at >>>> the documentation >>>>(https://docs.ceph.com/en/latest/rados/operations/erasure-code-lrc/)[90,113,2147483647,103,64,147,164,177,2147483647,133,58,28,8,32,2147483647]>>>> but I have some questions if someone has experience/expertisewith >>>> this LRC plugin. >>>>>>>> I tried to create a rule for using 5 OSDs per datacenter(15 in>>>> total), with 3 (9 in total) being data chunks and othersbeing >>>> coding chunks. For this, based of my understanding of examples, I >>>> used k=9, m=3, l=4. Is it right? Is this configuration equivalent,>>>> in terms of redundancy, to a jerasure configuration with k=9,m=6? >>>> >>>> The resulting rule, which looks correct to me, is: >>>> >>>> -------- >>>> >>>> { >>>> "rule_id": 6, >>>> "rule_name": "test_lrc_2", >>>> "ruleset": 6, >>>> "type": 3, >>>> "min_size": 3, >>>> "max_size": 15, >>>> "steps": [ >>>> { >>>> "op": "set_chooseleaf_tries", >>>> "num": 5 >>>> }, >>>> { >>>> "op": "set_choose_tries", >>>> "num": 100 >>>> }, >>>> { >>>> "op": "take", >>>> "item": -4, >>>> "item_name": "default~hdd" >>>> }, >>>> { >>>> "op": "choose_indep", >>>> "num": 3, >>>> "type": "datacenter" >>>> }, >>>> { >>>> "op": "chooseleaf_indep", >>>> "num": 5, >>>> "type": "host" >>>> }, >>>> { >>>> "op": "emit" >>>> } >>>> ] >>>> } >>>> >>>> ------------ >>>>>>>> Unfortunately, it doesn't work as expected: a pool createdwith>>>> this rule ends up with its pages active+undersize, which is >>>> unexpected for me. Looking at 'ceph health detail` output,I see >>>> for each page something like: >>>> >>>> pg 52.14 is stuck undersized for 27m, current state >>>> active+undersized, last acting >>>>>>>>>>>> For each PG, there is 3 '2147483647' entries and I guess itis the>>>> reason of the problem. What are these entries about? Clearlyit is>>>> not OSD entries... Looks like a negative number, -1, which interms>>>> of crushmap ID is the crushmap root (named "default" in our>>>> configuration). Any trivial mistake I would have made? >>>>>>>> Thanks in advance for any help or for sharing any successful>>>> configuration? >>>> >>>> Best regards, >>>> >>>> Michel >>>> _______________________________________________ >>>> ceph-users mailing list -- ceph-users@xxxxxxx>>>> To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
