Hi, A simple fsck eats the same amount of memory Cluster usage: rbd with a bit of rgw Here is the ceph df detail All OSDs are single rusty devices On 4/2/20 2:19 PM, Igor Fedotov wrote:

And high memory usage is present for quick-fix after conversion as well, isn't it? The same tens of GBs? On 4/2/2020 4:13 PM, Jack wrote: > (fsck / quick-fix, same story) > > On 4/2/20 3:12 PM, Jack wrote: >> Hi, >> >> A simple fsck eats the same amount of memory >> >> Cluster usage: rbd with a bit of rgw >> >> Here is the ceph df detail >> All OSDs are single rusty devices >> >> On 4/2/20 2:19 PM, Igor Fedotov wrote: >>> Hi Jack, >>> >>> could you please try the following - stop one of already converted OSDs >>> and do a quick-fix/fsck/repair against it using ceph_bluestore_tool: >>> >>> ceph-bluestore-tool --path <path to osd> --command >>> quick-fix|fsck|repair >>> >>> Does it cause similar memory usage? >>> >>> You can stop experimenting if quick-fix reproduces the issue. >>> >>> >>> Also could you please describe your cluster and its usage a bit: what's >>> the usage: rgw/rbd/cephfs? If possible - please share 'ceph df detail' >>> output, do you have standalone DB volume at SSD/NVMe? >>> >>> Thanks, >>> >>> Igor >>> >>> >>> On 4/1/2020 6:28 PM, Jack wrote: >>>> Hi, >>>> >>>> As the upgrade documentation tells: >>>>> Note that the first time each OSD starts, it will do a format >>>>> conversion to improve the accounting for “omap” data. This may >>>>> take a few minutes to as much as a few hours (for an HDD with lots >>>>> of omap data). You can disable this automatic conversion with: >>>> What the documentation does not say is that this process takes a >>>> lot of >>>> memory >>>> >>>> I am upgrading a rusty cluster from Nautilus, you can check out the >>>> ram >>>> consumption as attachment >>>> >>>> First, we have a 3TB osd conversion: it tooks ~15min, and 19GB of >>>> memory >>>> >>>> Then, we have a larger 6TB osd conversion: it tooks more than 2 hours, >>>> and 35GB of memory >>>> >>>> Finally, you have the largest 10TB osd: only 1H15, but 52GB of memory >>>> >>>> >>>> _______________________________________________ >>>> ceph-users mailing list -- ceph-users(a)ceph.io >>>> To unsubscribe send an email to ceph-users-leave(a)ceph.io

And may I have the output for: ceph daemon osd.N calc_objectstore_db_histogram This will collect some stats on record types in OSD's DB. On 4/2/2020 4:13 PM, Jack wrote: > (fsck / quick-fix, same story) > > On 4/2/20 3:12 PM, Jack wrote: >> Hi, >> >> A simple fsck eats the same amount of memory >> >> Cluster usage: rbd with a bit of rgw >> >> Here is the ceph df detail >> All OSDs are single rusty devices >> >> On 4/2/20 2:19 PM, Igor Fedotov wrote: >>> Hi Jack, >>> >>> could you please try the following - stop one of already converted OSDs >>> and do a quick-fix/fsck/repair against it using ceph_bluestore_tool: >>> >>> ceph-bluestore-tool --path <path to osd> --command >>> quick-fix|fsck|repair >>> >>> Does it cause similar memory usage? >>> >>> You can stop experimenting if quick-fix reproduces the issue. >>> >>> >>> Also could you please describe your cluster and its usage a bit: what's >>> the usage: rgw/rbd/cephfs? If possible - please share 'ceph df detail' >>> output, do you have standalone DB volume at SSD/NVMe? >>> >>> Thanks, >>> >>> Igor >>> >>> >>> On 4/1/2020 6:28 PM, Jack wrote: >>>> Hi, >>>> >>>> As the upgrade documentation tells: >>>>> Note that the first time each OSD starts, it will do a format >>>>> conversion to improve the accounting for “omap” data. This may >>>>> take a few minutes to as much as a few hours (for an HDD with lots >>>>> of omap data). You can disable this automatic conversion with: >>>> What the documentation does not say is that this process takes a >>>> lot of >>>> memory >>>> >>>> I am upgrading a rusty cluster from Nautilus, you can check out the >>>> ram >>>> consumption as attachment >>>> >>>> First, we have a 3TB osd conversion: it tooks ~15min, and 19GB of >>>> memory >>>> >>>> Then, we have a larger 6TB osd conversion: it tooks more than 2 hours, >>>> and 35GB of memory >>>> >>>> Finally, you have the largest 10TB osd: only 1H15, but 52GB of memory >>>> >>>> >>>> _______________________________________________ >>>> ceph-users mailing list -- ceph-users(a)ceph.io >>>> To unsubscribe send an email to ceph-users-leave(a)ceph.io

So this OSD has 32M of shared blobs and fsck loads them all into memory while processing. Hence the RAM consumption. I'm afraid there is no simple way to fix that, will create a ticket though. And a side question: 1) Do you use erasure coding and/or compression for rbd pool? These stats look suspicious POOL                        ID  STORED   (DATA)   (OMAP)   OBJECTS USED     (DATA)   (OMAP)   %USED  MAX AVAIL  QUOTA OBJECTS  QUOTA BYTES  DIRTY    USED COMPR  UNDER COMPR rbd                          1  245 TiB  245 TiB  9.0 MiB   50.26M 151 TiB  151 TiB  9.0 MiB  90.03     12 TiB  N/A N/A           50.26M      35 TiB      144 TiB Stored - 245 TiB, Used - 151 TiB Can't imagine any explanation other than applied compression. Thanks, Igor On 4/2/2020 5:59 PM, Jack wrote: > Here it is > > On 4/2/20 3:48 PM, Igor Fedotov wrote: >> And may I have the output for: >> >> ceph daemon osd.N calc_objectstore_db_histogram >> >> This will collect some stats on record types in OSD's DB. >> >> >> On 4/2/2020 4:13 PM, Jack wrote: >>> (fsck / quick-fix, same story) >>> >>> On 4/2/20 3:12 PM, Jack wrote: >>>> Hi, >>>> >>>> A simple fsck eats the same amount of memory >>>> >>>> Cluster usage: rbd with a bit of rgw >>>> >>>> Here is the ceph df detail >>>> All OSDs are single rusty devices >>>> >>>> On 4/2/20 2:19 PM, Igor Fedotov wrote: >>>>> Hi Jack, >>>>> >>>>> could you please try the following - stop one of already converted >>>>> OSDs >>>>> and do a quick-fix/fsck/repair against it using ceph_bluestore_tool: >>>>> >>>>> ceph-bluestore-tool --path <path to osd> --command >>>>> quick-fix|fsck|repair >>>>> >>>>> Does it cause similar memory usage? >>>>> >>>>> You can stop experimenting if quick-fix reproduces the issue. >>>>> >>>>> >>>>> Also could you please describe your cluster and its usage a bit: >>>>> what's >>>>> the usage: rgw/rbd/cephfs? If possible - please share 'ceph df >>>>> detail' >>>>> output, do you have standalone DB volume at SSD/NVMe? >>>>> >>>>> Thanks, >>>>> >>>>> Igor >>>>> >>>>> >>>>> On 4/1/2020 6:28 PM, Jack wrote: >>>>>> Hi, >>>>>> >>>>>> As the upgrade documentation tells: >>>>>>> Note that the first time each OSD starts, it will do a format >>>>>>> conversion to improve the accounting for “omap” data. This may >>>>>>> take a few minutes to as much as a few hours (for an HDD with lots >>>>>>> of omap data). You can disable this automatic conversion with: >>>>>> What the documentation does not say is that this process takes a >>>>>> lot of >>>>>> memory >>>>>> >>>>>> I am upgrading a rusty cluster from Nautilus, you can check out the >>>>>> ram >>>>>> consumption as attachment >>>>>> >>>>>> First, we have a 3TB osd conversion: it tooks ~15min, and 19GB of >>>>>> memory >>>>>> >>>>>> Then, we have a larger 6TB osd conversion: it tooks more than 2 >>>>>> hours, >>>>>> and 35GB of memory >>>>>> >>>>>> Finally, you have the largest 10TB osd: only 1H15, but 52GB of >>>>>> memory >>>>>> >>>>>> >>>>>> _______________________________________________ >>>>>> ceph-users mailing list -- ceph-users(a)ceph.io >>>>>> To unsubscribe send an email to ceph-users-leave(a)ceph.io

Thanks, Jack. One more question please - what's the actual maximum memory consumption for this specific OSD during fsck? And is it backed by 3, 6 or 10 TB  drive ? Regards, Igor On 4/2/2020 7:15 PM, Jack wrote: > I do compress: > root@backup2:~# ceph daemon osd.0 config show | grep > bluestore_compression >      "bluestore_compression_algorithm": "snappy", >      "bluestore_compression_max_blob_size": "0", >      "bluestore_compression_max_blob_size_hdd": "524288", >      "bluestore_compression_max_blob_size_ssd": "65536", >      "bluestore_compression_min_blob_size": "0", >      "bluestore_compression_min_blob_size_hdd": "8192", >      "bluestore_compression_min_blob_size_ssd": "8192", >      "bluestore_compression_mode": "force", >      "bluestore_compression_required_ratio": "0.955000", > > I will deal with the memory consumption > After all, it just require more time (starting OSD one by one), and it > still fits in my main memory > > Thank you for checkout out the issue > > > On 4/2/20 5:28 PM, Igor Fedotov wrote: >> So this OSD has 32M of shared blobs and fsck loads them all into memory >> while processing. Hence the RAM consumption. >> >> >> I'm afraid there is no simple way to fix that, will create a ticket >> though. >> >> >> And a side question: >> >> 1) Do you use erasure coding and/or compression for rbd pool? >> >> These stats look suspicious >> >> POOL                        ID  STORED   (DATA)   (OMAP)   OBJECTS >> USED     (DATA)   (OMAP)   %USED  MAX AVAIL  QUOTA OBJECTS  QUOTA BYTES >> DIRTY    USED COMPR  UNDER COMPR >> rbd                          1  245 TiB  245 TiB  9.0 MiB   50.26M 151 >> TiB  151 TiB  9.0 MiB  90.03     12 TiB  N/A N/A           50.26M >> 35 TiB      144 TiB >> >> Stored - 245 TiB, Used - 151 TiB >> >> Can't imagine any explanation other than applied compression. >> >> >> Thanks, >> >> Igor >> >> >> >> On 4/2/2020 5:59 PM, Jack wrote: >>> Here it is >>> >>> On 4/2/20 3:48 PM, Igor Fedotov wrote: >>>> And may I have the output for: >>>> >>>> ceph daemon osd.N calc_objectstore_db_histogram >>>> >>>> This will collect some stats on record types in OSD's DB. >>>> >>>> >>>> On 4/2/2020 4:13 PM, Jack wrote: >>>>> (fsck / quick-fix, same story) >>>>> >>>>> On 4/2/20 3:12 PM, Jack wrote: >>>>>> Hi, >>>>>> >>>>>> A simple fsck eats the same amount of memory >>>>>> >>>>>> Cluster usage: rbd with a bit of rgw >>>>>> >>>>>> Here is the ceph df detail >>>>>> All OSDs are single rusty devices >>>>>> >>>>>> On 4/2/20 2:19 PM, Igor Fedotov wrote: >>>>>>> Hi Jack, >>>>>>> >>>>>>> could you please try the following - stop one of already converted >>>>>>> OSDs >>>>>>> and do a quick-fix/fsck/repair against it using >>>>>>> ceph_bluestore_tool: >>>>>>> >>>>>>> ceph-bluestore-tool --path <path to osd> --command >>>>>>> quick-fix|fsck|repair >>>>>>> >>>>>>> Does it cause similar memory usage? >>>>>>> >>>>>>> You can stop experimenting if quick-fix reproduces the issue. >>>>>>> >>>>>>> >>>>>>> Also could you please describe your cluster and its usage a bit: >>>>>>> what's >>>>>>> the usage: rgw/rbd/cephfs? If possible - please share 'ceph df >>>>>>> detail' >>>>>>> output, do you have standalone DB volume at SSD/NVMe? >>>>>>> >>>>>>> Thanks, >>>>>>> >>>>>>> Igor >>>>>>> >>>>>>> >>>>>>> On 4/1/2020 6:28 PM, Jack wrote: >>>>>>>> Hi, >>>>>>>> >>>>>>>> As the upgrade documentation tells: >>>>>>>>> Note that the first time each OSD starts, it will do a format >>>>>>>>> conversion to improve the accounting for “omap” data. This may >>>>>>>>> take a few minutes to as much as a few hours (for an HDD with >>>>>>>>> lots >>>>>>>>> of omap data). You can disable this automatic conversion with: >>>>>>>> What the documentation does not say is that this process takes a >>>>>>>> lot of >>>>>>>> memory >>>>>>>> >>>>>>>> I am upgrading a rusty cluster from Nautilus, you can check out >>>>>>>> the >>>>>>>> ram >>>>>>>> consumption as attachment >>>>>>>> >>>>>>>> First, we have a 3TB osd conversion: it tooks ~15min, and 19GB of >>>>>>>> memory >>>>>>>> >>>>>>>> Then, we have a larger 6TB osd conversion: it tooks more than 2 >>>>>>>> hours, >>>>>>>> and 35GB of memory >>>>>>>> >>>>>>>> Finally, you have the largest 10TB osd: only 1H15, but 52GB of >>>>>>>> memory >>>>>>>> >>>>>>>> >>>>>>>> _______________________________________________ >>>>>>>> ceph-users mailing list -- ceph-users(a)ceph.io >>>>>>>> To unsubscribe send an email to ceph-users-leave(a)ceph.io