> On Apr 28, 2017, at 2:08 PM, Trond Myklebust <trondmy@xxxxxxxxxxxxxxx> wrote: > > On Fri, 2017-04-28 at 10:45 -0700, Chuck Lever wrote: >>> On Apr 28, 2017, at 10:25 AM, Trond Myklebust <trond.myklebust@prim >>> arydata.com> wrote: >>> >>> In the spirit of experimentation, I've put together a set of >>> patches >>> that implement setting up multiple TCP connections to the server. >>> The connections all go to the same server IP address, so do not >>> provide support for multiple IP addresses (which I believe is >>> something Andy Adamson is working on). >>> >>> The feature is only enabled for NFSv4.1 and NFSv4.2 for now; I >>> don't >>> feel comfortable subjecting NFSv3/v4 replay caches to this >>> treatment yet. It relies on the mount option "nconnect" to specify >>> the number of connections to st up. So you can do something like >>> 聽'mount -t nfs -overs=4.1,nconnect=8 foo:/bar /mnt' >>> to set up 8 TCP connections to server 'foo'. >> >> IMO this setting should eventually be set dynamically by the >> client, or should be global (eg., a module parameter). > > There is an argument for making it a per-server value (which is what > this patchset does). It allows the admin a certain control to limit the > number of connections to specific servers that are need to serve larger > numbers of clients. However I'm open to counter arguments. I've no > strong opinions yet. Like direct I/O, this kind of setting could allow a single client to DoS a server. One additional concern might be how to deal with servers who have no more ability to accept connections during certain periods, but are able to support a lot of connections at other times. >> Since mount points to the same server share the same transport, >> what happens if you specify a different "nconnect" setting on >> two mount points to the same server? > > Currently, the first one wins. > >> What will the client do if there are not enough resources >> (eg source ports) to create that many? Or is this an "up to N" >> kind of setting? I can imagine a big client having to reduce >> the number of connections to each server to help it scale in >> number of server connections. > > There is an arbitrary (compile time) limit of 16. The use of the > SO_REUSEPORT socket option ensures that we should almost always be able > to satisfy that number of source ports, since they can be shared with > connections to other servers. FWIW, Solaris limits this setting to 8. I think past that value, there is only incremental and diminishing gain. That could be apples to pears, though. I'm not aware of a mount option, but there might be a system tunable that controls this setting on each client. >> Other storage protocols have a mechanism for determining how >> transport connections are provisioned: One connection per >> CPU core (or one CPU per NUMA node) on the client. This gives >> a clear way to decide which connection to use for each RPC, >> and guarantees the reply will arrive at the same compute >> domain that sent the call. > > Can we perhaps lay out a case for which mechanisms are useful as far as > hardware is concerned? I understand the socket code is already > affinitised to CPU caches, so that one's easy. I'm less familiar with > the various features of the underlying offloaded NICs and how they tend > to react when you add/subtract TCP connections. Well, the optimal number of connections varies depending on the NIC hardware design. I don't think there's a hard-and-fast rule, but typically the server-class NICs have multiple DMA engines and multiple cores. Thus they benefit from having multiple sockets, up to a point. Smaller clients would have a handful of cores, a single memory hierarchy, and one NIC. I would guess optimizing for the NIC (or server) would be best in that case. I'd bet two connections would be a very good default. For large clients, a connection per NUMA node makes sense. This keeps the amount of cross-node memory traffic to a minimum, which improves system scalability. The issues with "socket per CPU core" are: there can be a lot of cores, and it might be wasteful to open that many sockets to each NFS server; and what do you do with a socket when a CPU core is taken offline? >> And of course: RPC-over-RDMA really loves this kind of feature >> (multiple connections between same IP tuples) to spread the >> workload over multiple QPs. There isn't anything special needed >> for RDMA, I hope, but I'll have a look at the SUNRPC pieces. > > I haven't yet enabled it for RPC/RDMA, but I imagine you can help out > if you find it useful (as you appear to do). I can give the patch set a try this week. I haven't seen any thing that would exclude proto=rdma from playing in this sandbox. >> Thanks for posting, I'm looking forward to seeing this >> capability in the Linux client. >> >> >>> Anyhow, feel free to test and give me feedback as to whether or not >>> this helps performance on your system. >>> >>> Trond Myklebust (5): >>> 聽SUNRPC: Allow creation of RPC clients with multiple connections >>> 聽NFS: Add a mount option to specify number of TCP connections to >>> use >>> 聽NFSv4: Allow multiple connections to NFSv4.x (x>0) servers >>> 聽pNFS: Allow multiple connections to the DS >>> 聽NFS: Display the "nconnect" mount option if it is set. >>> >>> fs/nfs/client.c聽聽聽聽聽聽聽聽聽聽聽聽聽|聽聽2 ++ >>> fs/nfs/internal.h聽聽聽聽聽聽聽聽聽聽聽|聽聽2 ++ >>> fs/nfs/nfs3client.c聽聽聽聽聽聽聽聽聽|聽聽3 +++ >>> fs/nfs/nfs4client.c聽聽聽聽聽聽聽聽聽| 13 +++++++++++-- >>> fs/nfs/super.c聽聽聽聽聽聽聽聽聽聽聽聽聽聽| 12 ++++++++++++ >>> include/linux/nfs_fs_sb.h聽聽聽|聽聽1 + >>> include/linux/sunrpc/clnt.h |聽聽1 + >>> net/sunrpc/clnt.c聽聽聽聽聽聽聽聽聽聽聽| 17 ++++++++++++++++- >>> net/sunrpc/xprtmultipath.c聽聽|聽聽3 +-- >>> 9 files changed, 49 insertions(+), 5 deletions(-) >>> >>> --聽 >>> 2.9.3 >>> >>> -- >>> To unsubscribe from this list: send the line "unsubscribe linux- >>> nfs" in >>> the body of a message to majordomo@xxxxxxxxxxxxxxx >>> More majordomo info at聽聽http://vger.kernel.org/majordomo-info.html >> >> -- >> Chuck Lever >> >> >> > -- > Trond Myklebust > Linux NFS client maintainer, PrimaryData > trond.myklebust@xxxxxxxxxxxxxxx > �N嫥叉靣笡y氊b瞂千v豝�)藓{.n�+壏{睗�"炟^n噐■��侂h櫒璀�&Ⅷ�瓽珴閔��(殠娸"濟���m��飦赇z罐枈帼f"穐殘坢 -- Chuck Lever -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html