dev: add support for multi gem5 runs

Multi gem5 is an extension to gem5 to enable parallel simulation of a
distributed system (e.g. simulation of a pool of machines
connected by Ethernet links). A multi gem5 run consists of seperate gem5
processes running in parallel (potentially on different hosts/slots on
a cluster). Each gem5 process executes the simulation of a component of the
simulated distributed system (e.g. a multi-core board with an Ethernet NIC).

The patch implements the "distributed" Ethernet link device
(dev/src/multi_etherlink.[hh.cc]). This device will send/receive
(simulated) Ethernet packets to/from peer gem5 processes. The interface
to talk to the peer gem5 processes is defined in dev/src/multi_iface.hh and
in tcp_iface.hh.

There is also a central message server process (util/multi/tcp_server.[hh,cc])
which acts like an Ethernet switch and transfers messages among the gem5 peers.

A multi gem5 simulations can be kicked off by the util/multi/gem5-multi.sh
wrapper script.

Checkpoints are supported by multi-gem5. The checkpoint must be
initiated by a single gem5 process. E.g., the gem5 process with rank 0
can take a checkpoint from the bootscript just before it invokes
'mpirun' to launch an MPI test. The message server process will notify
all the other peer gem5 processes and make them take a checkpoint, too
(after completing a global synchronisation to ensure that there are no
inflight messages among gem5).

1 files changed, 492 insertions, 0 deletions

diff --git a/src/dev/multi_iface.hh b/src/dev/multi_iface.hh
new file mode 100644
index 000000000..0e4859ecd
--- /dev/null
+++ b/src/dev/multi_iface.hh
@@ -0,0 +1,492 @@
+/*
+ * Copyright (c) 2015 ARM Limited
+ * All rights reserved
+ *
+ * The license below extends only to copyright in the software and shall
+ * not be construed as granting a license to any other intellectual
+ * property including but not limited to intellectual property relating
+ * to a hardware implementation of the functionality of the software
+ * licensed hereunder.  You may use the software subject to the license
+ * terms below provided that you ensure that this notice is replicated
+ * unmodified and in its entirety in all distributions of the software,
+ * modified or unmodified, in source code or in binary form.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met: redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer;
+ * redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution;
+ * neither the name of the copyright holders nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * Authors: Gabor Dozsa
+ */
+
+/* @file
+ * The interface class for multi gem5 simulations.
+ *
+ * Multi gem5 is an extension to gem5 to enable parallel simulation of a
+ * distributed system (e.g. simulation of a pool of machines
+ * connected by Ethernet links). A multi gem5 run consists of seperate gem5
+ * processes running in parallel. Each gem5 process executes
+ * the simulation of a component of the simulated distributed system.
+ * (An example component can be a multi-core board with an Ethernet NIC.)
+ * The MultiIface class below provides services to transfer data and
+ * control messages among the gem5 processes. The main such services are
+ * as follows.
+ *
+ * 1. Send a data packet coming from a simulated Ethernet link. The packet
+ * will be transferred to (all) the target(s) gem5 processes. The send
+ * operation is always performed by the simulation thread, i.e. the gem5
+ * thread that is processing the event queue associated with the simulated
+ * Ethernet link.
+ *
+ * 2. Spawn a receiver thread to process messages coming in from the
+ * from other gem5 processes. Each simulated Ethernet link has its own
+ * associated receiver thread. The receiver thread saves the incoming packet
+ * and schedule an appropriate receive event in the event queue.
+ *
+ * 3. Schedule a global barrier event periodically to keep the gem5
+ * processes in sync.
+ * Periodic barrier event to keep peer gem5 processes in sync. The basic idea
+ * is that no gem5 process can go ahead further than the simulated link
+ * transmission delay to ensure that a corresponding receive event can always
+ * be scheduled for any message coming in from a peer gem5 process.
+ *
+ *
+ *
+ * This interface is an abstract class (sendRaw() and recvRaw()
+ * methods are pure virtual). It can work with various low level
+ * send/receive service implementations (e.g. TCP/IP, MPI,...). A TCP
+ * stream socket version is implemented in dev/src/tcp_iface.[hh,cc].
+ */
+#ifndef __DEV_MULTI_IFACE_HH__
+#define __DEV_MULTI_IFACE_HH__
+
+#include <array>
+#include <mutex>
+#include <queue>
+#include <thread>
+#include <utility>
+
+#include "dev/etherpkt.hh"
+#include "dev/multi_packet.hh"
+#include "sim/core.hh"
+#include "sim/drain.hh"
+#include "sim/global_event.hh"
+
+class EventManager;
+
+/**
+ * The interface class to talk to peer gem5 processes.
+ */
+class MultiIface : public Drainable
+{
+  public:
+    /*!
+     * The possible reasons a multi sync among gem5 peers is needed for.
+     */
+    enum
+    class SyncTrigger {
+        periodic, /*!< Regular periodic sync. This can be interrupted by a
+                   checkpoint sync request */
+        ckpt,     /*!< sync before taking a checkpoint */
+        atomic    /*!< sync that cannot be interrupted (e.g. sync at startup) */
+    };
+
+  private:
+    typedef MultiHeaderPkt::MsgType MsgType;
+
+    /** Sync State-Machine
+     \dot
+     digraph Sync {
+     node [shape=box, fontsize=10];
+     idle -> busy
+     [ label="new trigger\n by run()" fontsize=8 ];
+     busy -> busy
+     [ label="new message by progress():\n(msg == SyncAck &&\nwaitNum > 1) || \n(msg==CkptSyncReq &&\ntrigger == ckpt)" fontsize=8 ];
+     busy -> idle
+     [ label="new message by progress():\n(msg == SyncAck &&\nwaitNum == 1)" fontsize=8 ];
+     busy -> interrupted
+     [ label="new message by progress():\n(msg == CkptSyncReq &&\ntrigger == periodic)" fontsize=8 ];
+     idle -> asyncCkpt
+     [ label="new message by progress():\nmsg == CkptSyncReq" fontsize=8 ];
+     asyncCkpt -> asyncCkpt
+     [ label="new message by progress():\nmsg == CkptSyncReq" fontsize=8 ];
+     asyncCkpt -> busy
+     [ label="new trigger by run():\ntrigger == ckpt" fontsize=8 ];
+     asyncCkpt -> idle
+     [ label="new trigger by run():\n(trigger == periodic &&\nwaitNum == 0) " fontsize=8 ];
+     asyncCkpt -> interrupted
+     [ label="new trigger by run():\n(trigger == periodic &&\nwaitNum > 0) " fontsize=8 ];
+     interrupted -> interrupted
+     [ label="new message by progress():\n(msg == CkptSyncReq &&\nwaitNum > 1)" fontsize=8 ];
+     interrupted -> idle
+     [ label="new message by progress():\n(msg == CkptSyncReq &&\nwaitNum == 1)" fontsize=8 ];
+     }
+     \enddot
+     */
+    /** @class Sync
+     * This class implements global sync operations among gem5 peer processes.
+     *
+     * @note This class is used as a singleton object (shared by all MultiIface
+     * objects).
+     */
+    class Sync
+    {
+      private:
+        /*!
+         * Internal state of the sync singleton object.
+         */
+        enum class SyncState {
+            busy,        /*!< There is an on-going sync. */
+            interrupted, /*!< An on-going periodic sync was interrupted. */
+            asyncCkpt,   /*!< A checkpoint (sim_exit) is already scheduled */
+            idle         /*!< There is no active sync. */
+        };
+        /**
+         * The lock to protect access to the MultiSync object.
+         */
+        std::mutex lock;
+        /**
+         * Condition variable for the simulation thread to wait on
+         * until all receiver threads completes the current global
+         * synchronisation.
+         */
+        std::condition_variable cv;
+        /**
+         * Number of receiver threads that not yet completed the current global
+         * synchronisation.
+         */
+        unsigned waitNum;
+        /**
+         * The trigger for the most recent sync.
+         */
+        SyncTrigger trigger;
+        /**
+         * Map sync triggers to request messages.
+         */
+        std::array<MsgType, 3> triggerToMsg = {{
+                MsgType::cmdPeriodicSyncReq,
+                MsgType::cmdCkptSyncReq,
+                MsgType::cmdAtomicSyncReq
+            }};
+
+        /**
+         * Current sync state.
+         */
+        SyncState state;
+
+      public:
+        /**
+         *  Core method to perform a full multi sync.
+         *
+         * @param t Sync trigger.
+         * @param sync_tick The tick the sync was expected to happen at.
+         * @return true if the sync completed, false if it was interrupted.
+         *
+         * @note In case of an interrupted periodic sync, sync_tick can be less
+         * than curTick() when we resume (i.e. re-run) it
+         */
+        bool run(SyncTrigger t, Tick sync_tick);
+        /**
+         * Callback when the receiver thread gets a sync message.
+         */
+        void progress(MsgType m);
+
+        Sync() : waitNum(0), state(SyncState::idle) {}
+        ~Sync() {}
+    };
+
+
+    /**
+     * The global event to schedule peridic multi sync. It is used as a
+     * singleton object.
+     *
+     * The periodic synchronisation works as follows.
+     * 1. A MultisyncEvent is scheduled as a global event when startup() is
+     * called.
+     * 2. The progress() method of the MultisyncEvent initiates a new barrier
+     * for each simulated Ethernet links.
+     * 3. Simulation thread(s) then waits until all receiver threads
+     * completes the ongoing barrier. The global sync event is done.
+     */
+    class SyncEvent : public GlobalSyncEvent
+    {
+      public:
+        /**
+         * Flag to indicate that the most recent periodic sync was interrupted
+         * (by a checkpoint request).
+         */
+        bool interrupted;
+        /**
+         * The tick when the most recent periodic synchronisation was scheduled
+         * at.
+         */
+        Tick scheduledAt;
+        /**
+         * Flag to indicate an on-going drain cycle.
+         */
+         bool isDraining;
+
+      public:
+        /**
+         * Only the firstly instanstiated MultiIface object will
+         * call this constructor.
+         */
+        SyncEvent() : GlobalSyncEvent(Default_Pri, 0), interrupted(false),
+                      scheduledAt(0), isDraining(false) {}
+
+        ~SyncEvent() { assert (scheduled() == false); }
+        /**
+         * Schedule the first periodic sync event.
+         *
+         * @param start Start tick for multi synchronisation
+         * @param repeat Frequency of multi synchronisation
+         *
+         */
+        void start(Tick start, Tick repeat);
+        /**
+         * Reschedule (if necessary) the periodic sync event.
+         *
+         * @param start Start tick for multi synchronisation
+         * @param repeat Frequency of multi synchronisation
+         *
+         * @note Useful if we have multiple MultiIface objects with
+         * different 'start' and 'repeat' values for global sync.
+         */
+        void adjust(Tick start, Tick repeat);
+        /**
+         * This is a global event so process() will be called by each
+         * simulation threads. (See further comments in the .cc file.)
+         */
+        void process() M5_ATTR_OVERRIDE;
+        /**
+         * Schedule periodic sync when resuming from a checkpoint.
+         */
+        void resume();
+
+        void serialize(const std::string &base, CheckpointOut &cp) const;
+        void unserialize(const std::string &base, CheckpointIn &cp);
+    };
+
+    /**
+     * The receive thread needs to store the packet pointer and the computed
+     * receive tick for each incoming data packet. This information is used
+     * by the simulation thread when it processes the corresponding receive
+     * event. (See more comments at the implemetation of the recvThreadFunc()
+     * and RecvPacketIn() methods.)
+     */
+    typedef std::pair<EthPacketPtr, Tick> RecvInfo;
+
+    /**
+     * Comparison predicate for RecvInfo, needed by the recvQueue.
+     */
+    struct RecvInfoCompare {
+        bool operator()(const RecvInfo &lhs, const RecvInfo &rhs)
+        {
+            return lhs.second > rhs.second;
+        }
+    };
+
+    /**
+     * Customized priority queue used to store incoming data packets info by
+     * the receiver thread. We need to expose the underlying container to
+     * enable iterator access for serializing.
+     */
+    class RecvQueue : public std::priority_queue<RecvInfo,
+                                                 std::vector<RecvInfo>,
+                                                 RecvInfoCompare>
+    {
+      public:
+        std::vector<RecvInfo> &impl() { return c; }
+        const std::vector<RecvInfo> &impl() const { return c; }
+    };
+
+    /*
+     * The priority queue to store RecvInfo items ordered by receive ticks.
+     */
+    RecvQueue recvQueue;
+    /**
+     * The singleton Sync object to perform multi synchronisation.
+     */
+    static Sync *sync;
+    /**
+     * The singleton SyncEvent object to schedule periodic multi sync.
+     */
+    static SyncEvent *syncEvent;
+    /**
+     * Tick to schedule the first multi sync event.
+     * This is just as optimization : we do not need any multi sync
+     * event until the simulated NIC is brought up by the OS.
+     */
+    Tick syncStart;
+    /**
+     * Frequency of multi sync events in ticks.
+     */
+    Tick syncRepeat;
+    /**
+     * Receiver thread pointer.
+     * Each MultiIface object must have exactly one receiver thread.
+     */
+    std::thread *recvThread;
+    /**
+     * The event manager associated with the MultiIface object.
+     */
+    EventManager *eventManager;
+
+    /**
+     * The receive done event for the simulated Ethernet link.
+     * It is scheduled by the receiver thread for each incoming data
+     * packet.
+     */
+    Event *recvDone;
+
+    /**
+     * The packet that belongs to the currently scheduled recvDone event.
+     */
+    EthPacketPtr scheduledRecvPacket;
+
+    /**
+     * The link delay in ticks for the simulated Ethernet link.
+     */
+    Tick linkDelay;
+
+    /**
+     * The rank of this process among the gem5 peers.
+     */
+    unsigned rank;
+    /**
+     * Total number of receiver threads (in this gem5 process).
+     * During the simulation it should be constant and equal to the
+     * number of MultiIface objects (i.e. simulated Ethernet
+     * links).
+     */
+    static unsigned recvThreadsNum;
+    /**
+     * The very first MultiIface object created becomes the master. We need
+     * a master to co-ordinate the global synchronisation.
+     */
+    static MultiIface *master;
+
+  protected:
+    /**
+     * Low level generic send routine.
+     * @param buf buffer that holds the data to send out
+     * @param length number of bytes to send
+     * @param dest_addr address of the target (simulated NIC). This may be
+     * used by a subclass for optimization (e.g. optimize broadcast)
+     */
+    virtual void sendRaw(void *buf,
+                         unsigned length,
+                         const MultiHeaderPkt::AddressType dest_addr) = 0;
+    /**
+     * Low level generic receive routine.
+     * @param buf the buffer to store the incoming message
+     * @param length buffer size (in bytes)
+     */
+    virtual bool recvRaw(void *buf, unsigned length) = 0;
+    /**
+     * Low level request for synchronisation among gem5 processes. Only one
+     * MultiIface object needs to call this (in each gem5 process) to trigger
+     * a multi sync.
+     *
+     * @param sync_req Sync request command.
+     * @param sync_tick The tick when sync is expected to happen in the sender.
+     */
+    virtual void syncRaw(MsgType sync_req, Tick sync_tick) = 0;
+    /**
+     * The function executed by a receiver thread.
+     */
+    void recvThreadFunc();
+    /**
+     * Receive a multi header packet. Called by the receiver thread.
+     * @param header the structure to store the incoming header packet.
+     * @return false if any error occured during the receive, true otherwise
+     *
+     * A header packet can carry a control command (e.g. 'barrier leave') or
+     * information about a data packet that is following the header packet
+     * back to back.
+     */
+    bool recvHeader(MultiHeaderPkt::Header &header);
+    /**
+     * Receive a data packet. Called by the receiver thread.
+     * @param data_header The packet descriptor for the expected incoming data
+     * packet.
+     */
+    void recvData(const MultiHeaderPkt::Header &data_header);
+
+  public:
+
+    /**
+     * ctor
+     * @param multi_rank Rank of this gem5 process within the multi run
+     * @param sync_start Start tick for multi synchronisation
+     * @param sync_repeat Frequency for multi synchronisation
+     * @param em The event manager associated with the simulated Ethernet link
+     */
+    MultiIface(unsigned multi_rank,
+               Tick sync_start,
+               Tick sync_repeat,
+               EventManager *em);
+
+    virtual ~MultiIface();
+    /**
+     * Send out an Ethernet packet.
+     * @param pkt The Ethernet packet to send.
+     * @param send_delay The delay in ticks for the send completion event.
+     */
+    void packetOut(EthPacketPtr pkt, Tick send_delay);
+    /**
+     * Fetch the next packet from the receive queue.
+     */
+    EthPacketPtr packetIn();
+
+    /**
+     * spawn the receiver thread.
+     * @param recv_done The receive done event associated with the simulated
+     * Ethernet link.
+     * @param link_delay The link delay for the simulated Ethernet link.
+     */
+    void spawnRecvThread(Event *recv_done,
+                         Tick link_delay);
+    /**
+     * Initialize the random number generator with a different seed in each
+     * peer gem5 process.
+     */
+    void initRandom();
+
+    DrainState drain() M5_ATTR_OVERRIDE;
+
+    /**
+     * Callback when draining is complete.
+     */
+    void drainDone();
+
+    /**
+     * Initialize the periodic synchronisation among peer gem5 processes.
+     */
+    void startPeriodicSync();
+
+    void serialize(const std::string &base, CheckpointOut &cp) const;
+    void unserialize(const std::string &base, CheckpointIn &cp);
+
+};
+
+
+#endif