P2P Design Specification

From Linux NFS

Latest revision as of 19:23, 18 January 2013

Overview

Peer-to-peer pNFS is designed to solve the "boot storm" problem that happens when several clients in a cluster boot and attempt to read the same set of files from a single NFS server all at the same time. This could overload the server's bandwidth, slowing down operations on most client machines. The idea behind p2p NFS is to allow clients to act as an adhoc read-only pNFS data server that serves files out of their data cache. This should spread out network usage across all machines, rather than focusing all activity on a single node. Server and desired DS machines will need to be modified but any pNFS-enabled client already has the code required to read from adhoc DSs.

Related Documents

• draft-myklebust-nfsv4-pnfs-backend-protocol-01.txt
• RFC 5661

Dependencies

This design needs the following from others:

Assumptions

• Workload with large number of read-only files
• Enable the following .config options for the pNFS client and pNFS ds machines:
  • CONFIG_NFS_V4_1
  • CONFIG_PNFS_FILE_LAYOUT
• Enable the following .config option for the pNFS ds machine:
  • CONFIG_NFS_P2P
• Enable the following .config options for the pNFS server and pNFS ds machines:
  • CONFIG_PNFSD
  • CONFIG_PNFSD_LOCAL_EXPORT
  • CONFIG_PNFSD_P2P
• Install pnfs-nfs-utils on the pNFS server
• Add "pnfs" to the export options of a local filesystem on the pNFS server
• pNFS DS should have nfsd running, but does not need to edit /etc/exports to share files
• pnfsd needs to add "pnfs" export option to /etc/exports
• pnfsd also needs to have "fsid=0" as an export option, otherwise the path walking code will trigger an early UNREGISTER_DS.

Design

REGISTER_DS

• Server
  • Only implemented REGISTER_DS_ALL
  • Create a new struct pnfs_p2p_client to store information about the adhoc DS:
    • p2p client stateid
    • netid
    • ip address
    • MDS identifier
  • Store structure as part of the nfs4_client
  • Encode p2p client stateid as reply to client
• Client
  • Send REGISTER_DS call as part of nfs4_remote_mount()
    • Use REGISTER_DS_ALL so server knows we'll cache everything
    • Generate MDS identifier using cl_cb_ident and a static u32 counter

UNREGISTER_DS

• Server
  • Check that the nfs4_client has an associated pnfs_p2p_client
  • Check that the nfs4_client is using the correct stateid
  • Free memory allocated for struct pnfs_p2p_client structure during REGISTER_DS
  • Free pnfs_p2p_po_stids associated with the DS
  • Set pnfs_p2p_client pointer in nfs4_client to NULL
• Client
  • Send UNREGISTER_DS as part of nfs4_destroy_server()

PROXY_OPEN

• Server
  • Introduce a pnfs_p2p_po_stid to track what DS the client was referred to
  • Strip MDS ID from the filehandle
  • Add stateid to list stored in the pnfs_p2p_client for the DS
  • Add stateid to list stored in the nfs4_client for the client
  • Initialize a callback workqueue structure for PROXY_REVOKE
• Client
  • Check if we have already called PROXY_OPEN for this (filehandle, stateid)
  • Check that we still have a delegation for the file
  • Use MDS identifier from filehandle to find the correct nfs_server structure
  • Use server to call an nfs4_proc_proxy_open()
    • Pass filehandle and read stateid
    • Use the compound: [SEQUENCE, PUTFH, PROXY_OPEN,GETFH] to look up the actual filehandle and get a proxy revoke stateid
  • Store both filehandles, read stateid and revoke stateid in a pnfs_po_state structure
    • Store this in the pnfs_layout_hdr
  • Pass resulting filehandle to nfs_delegation_find_inode() to find inode
  • Use d_find_any_alias() on the inode to find and return a dentry to the server

CB_PROXY_REVOKE

• Server
  • Call when client expires on server
  • Remove pnfs_p2p_po_stid from lists, but don't free until proxy_revoke_release()
• Client
  • Use the filehandle and stateid to find associated layout
  • Free that pnfs_po_stid

LAYOUTGET

• Server
  • Edit pnfs_lexp_layout_get()
  • Set device id field in the layout to the clientid of the machine acting as the DS
  • If we are not using p2p for the file, instead continue to return 1 as the devid
  • Encode a filehandle with the DSs MDS ID prepended in filelayout_encode_layout()

LAYOUTRETURN

• Server
  • Add to pnfs_lexp_layout_return()
  • Check nfs4_client for files opened on a DS
    • Send CB_PROXY_REVOKE
  • Also check the pnfs_p2p_client structure for files cached as a DS
    • Free these stateids directly
• Client
  • Free up pnfs_po_state stored in the pnfs_layout_hdr

GETDEVICEINFO

• Server
  • If we are given a device id of 1 continue using the non-p2p code
  • Edit pnfsd_lexp_get_device_info() to fill out pnfs_filelayout_devaddr structure with DS information
  • Translate deviceid back to clientid to look up the DS
  • Fill out netid and ip address information using data in the pnfs_p2p_client structure

PUTFH

• Server
  • If this is a p2p filehandle then skip some of the state checking stuff because we won't have a dentry until after calling PROXY_OPEN
  • Check if a filehandle is p2p by looking at the length (p2p: 36 bytes, normal: 28 bytes)

READ

• Server
  • Call into the NFS client module to perform PROXY_OPEN and return the associated dentry for p2p filehandles

OPEN

• Server
  • Introduce a vfs_find_any_mount() to look up any mount structure for a dentry
    • This is a hack, but we don't care which mount structure as long as we get the file data!

Other Notes

• free_p2p_po_stid()
  • Remove from lists first before either freeing or calling CB_PROXY_REVOKE to prevent accidental double frees
• DS expires on server
  • Treat as if the client had called unregister_ds()

Data Structures

Server

• p2p client information

struct pnfs_p2p_client {
       struct nfs4_stid p2p_stid;
       u64 p2p_mds_id;
       char *p2p_netid;
       char *p2p_addr;
       struct list_head p2p_ds_files;
};

• p2p proxy open stateid

struct pnfs_p2p_po_stid {
       struct nfs4_stid  po_stid;
       struct knfsd_fh   po_fh;
       struct list_head  po_ds_list;
       struct list_head  po_cl_list;
       struct nfsd4_callback po_cb;
};

Client

• NFSv4 Proxy Open

struct pnfs_po_state {
       nfs4_stateid  read_stateid;
       nfs4_stateid  revoke_stateid;
       struct nfs_fh fh;
       struct list_head list;
};

Compatibility

• Any v4.1 / pNFS enabled client should be able to make use of adhoc data servers already, and not need special p2p extensions.
• Clients wishing to act as a data server need CONFIG_NFS_P2P enabled
• Servers wishing to track adhoc DSs need CONFIG_PNFSD_P2P enabled

Documentation

• I can write a Documentation/filesystems/nfs/peer_to_peer.txt file to give a brief overview of how p2pNFS is supposed to work and how users can configure it.
• I can also copy the page to linux-nfs.org for "online documentation"

Feature Interaction Dependencies and Impacts

• nfsd <-> nfs
  • The machine acting as a pNFS DS needs to be running both the nfs server and the nfs client.
• Made changes to putfh
  • Check filehandle length since p2p filehandles are longer
  • Call the original version of the function if we are using a normal fh
• Made changes to nfsd4_read
  • Call original read function if this isn't p2p, call proxy open otherwise to get data from client
• nfsd_open needs to lookup mount structure without using an exportops structure for NFS
• filelayout_encode_layout needs to be able to encode p2p filehandles and normal filehandles
• pnfs_p2p_mark_fh increases filehandle size, server needs to know to use the mds id for bigger filehandles
• nfsd4_proc_compound needs to know if a filehandle is a p2p fh since the dentry will be looked up later for reads

Performance

• Keep a per-file LRU list of clients that currently have the file cached to avoid redirecting all p2p activity to the same client for that file.

Scalability

The hope is that p2p NFS scales to hundreds and thousands of clients better than straight pnfs does. This can be tested by comparing read times for files of varying sizes both with and without p2p enabled. A handful of DSs and a large number of clients should be used to get a feel for how this would work in a data center.

• An LRU list of clients should help load balance traffic to each DS
  • Make use of already existing nfs4_file->fi_delegations list, move a DSs delegation to the end when referring
• I take the state lock (global mutex) when accessing file or client state
• I created a p2p spinlock for accessing p2p state

Testing

• Basic proof-of-concept tests
  • 1 client, 1 DS, 1 server
  • Have DS and client rsync files from server
  • Maybe do a `git clone linux-src` instead?
  • Try exporting a /lib partition

• In-depth testing
  • NFSv4root with varying numbers of clients
    • NFSv4root doesn't work right now due to idmapping issues
  • More rsyncs / git clones with more clients

Open Issues

Approvals

Approvers

Reviewers