Wonderful, just as I expected. Do folks normally have several RGW running on individual machines with a load balancer at larger scales?
On Wed, Dec 21, 2016 at 8:22 AM, LOPEZ Jean-Charles <jelo...@redhat.com> wrote: > Hi Gerald, > > for the s3 and swift case, the clients are not accessing the ceph cluster. > They are s3 and swift clients and only discuss with the RGW over HTTP. The > RGW is the ceph client that does all the interaction with the ceph cluster. > > Best > JC > > On Dec 21, 2016, at 07:27, Gerald Spencer <ger.spenc...@gmail.com> wrote: > > I was under the impression that when a client talks to the cluster, it > grabs the osd map and computes the crush algorithm to determine where it > stores the object. Does the rgw server do this for clients? If I had 12 > clients all talking through one gateway, would that server have to pass all > of the objects from the clients to the cluster? > > > And 48 osd nodes, each with 12 x 6TB drives and a PCIe write journal. That > would be 576 osds in the cluster, with about 3.4PB raw... > > > On Tue, Dec 20, 2016 at 1:12 AM Wido den Hollander <w...@42on.com> wrote: > >> >> >> > Op 20 december 2016 om 3:24 schreef Gerald Spencer < >> ger.spenc...@gmail.com>: >> >> > >> >> > >> >> > Hello all, >> >> > >> >> > We're currently waiting on a delivery of equipment for a small 50TB >> proof >> >> > of concept cluster, and I've been lurking/learning a ton from you. >> Thanks >> >> > for how active everyone is. >> >> > >> >> > Question(s): >> >> > How does the raids gateway work exactly? >> >> >> >> The RGW doesn't do any RAID. It chunks up larger objects into smaller >> RADOS chunks. The first chunk is always 512k (IIRC) and then it chunks up >> into 4MB RADOS objects. >> >> >> >> > Does it introduce a single point of failure? >> >> >> >> It does if you deploy only one RGW. Always deploy multiple with >> loadbalancing in front. >> >> >> >> > Does all of the traffic go through the host running the rgw server? >> >> >> >> Yes it does. >> >> >> >> > >> >> > I just don't fully understand that side of things. As for architecture >> our >> >> > poc will have: >> >> > - 1 monitor >> >> > - 4 OSDs with 12 x 6TB drives, 1 x 800 PCIe journal >> >> > >> >> >> >> Underscaled machines, go for less disks per machine but more machines. >> More smaller machines works a lot better with Ceph then a few big machines. >> >> >> >> > I'd all goes as planned, this will scale up to: >> >> > - 3 monitors >> >> >> >> Always run with 3 MONs. Otherwise it is a serious SPOF. >> >> >> >> > - 48 osds >> >> > >> >> > This should give us enough storage (~1.2PB) wth enough throughput to >> handle >> >> > the data requirements of our machines to saturate our 100Gb link... >> >> > >> >> >> >> That won't happen with just 4 machines. Replica 3x taken into account is >> well. You will need a lot more machines to get the 100Gb link fully >> utilized. >> >> >> >> Wido >> >> >> >> > >> >> > >> >> > >> >> > >> >> > Cheers, >> >> > G >> >> > _______________________________________________ >> >> > ceph-users mailing list >> >> > ceph-users@lists.ceph.com >> >> > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >> >> _______________________________________________ > ceph-users mailing list > ceph-users@lists.ceph.com > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > > >
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
