Zoltan2_Directory.hpp

Go to the documentation of this file.

/*
 * @HEADER
 *
 * ***********************************************************************
 *
 *  Zoltan2 Directory for Load-balancing, Partitioning, Ordering and Coloring
 *                  Copyright 2012 Sandia Corporation
 *
 * Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 * the U.S. Government retains certain rights in this software.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are
 * met:
 *
 * 1. Redistributions of source code must retain the above copyright
 * notice, this list of conditions and the following disclaimer.
 *
 * 2. 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.
 *
 * 3. Neither the name of the Corporation nor the names of the
 * contributors may be used to endorse or promote products derived from
 * this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "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 SANDIA CORPORATION OR THE
 * 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.
 *
 * Questions? Contact Karen Devine  kddevin@sandia.gov
 *                    Erik Boman  egboman@sandia.gov
 *
 * ***********************************************************************
 *
 * @HEADER
 */

#ifndef ZOLTAN2_DIRECTORY_H_
#define ZOLTAN2_DIRECTORY_H_

#include <Teuchos_DefaultComm.hpp> // currently using Teuchos comm throughout

#ifndef HAVE_MPI
// support mpi serial - directory currently has a mix of Teuchos mpi commands
// and original MPI commands so as this gets better organized the serial support
// can be handled more naturally. Currently the tests call an MPI_Init for serial
// so this can all work. I've mostly added the serial support just to make
// debugging easier.
#include <mpi.h>
#endif

#include <Kokkos_UnorderedMap.hpp> // unordered map stores the local nodes

namespace Zoltan2 {

// The new Kokkos mode maps over the gid using unordered map
// Originally was thinking we don't want ptrs here and want the user data
// to be a natural type so it can cleanly support std::vector. However as this
// evolved I am not sure this is the best decision and can be more costly for
// memory. Now that all the unit testing is in place it would be easier to
// refactor this back to the original packed ptr style if necessary.
template <typename gid_t, typename lid_t, typename user_t>
class Zoltan2_Directory_Node {
  public:
    // TODO: This free is an outstanding issue - follows original format and
    // allows me to search the Kokkos::UnorderedMap by index and distinguish
    // between empty slots and filled. But I am not quite understanding this
    // and is there a way to search the filled slots without referring to the
    // keys.
    Zoltan2_Directory_Node() : free(1) {}
    int owner;       /* processor hosting global ID object    */
    int partition;   /* Optional data                         */
    int errcheck;    /* Error checking(inconsistent updates)  */
    lid_t lid;       /* lid value */
    user_t userData; /* user data */
    int free;        /* flag whether node is free or used     */
};

// TODO: These message structures should become MPI datatypes(KDD)
// Currently these are implemented as they were in the original zoltan code.
template <typename gid_t, typename lid_t>
class Zoltan2_DD_Update_Msg {  /* Only used by Zoltan_DD_Update()             */
  public:
   char lid_flag;              /* indicates if LID data are present           */
   char user_flag;             /* indicates if USER data are present          */
   char partition_flag;        /* indicates if optional partition data        */
   int owner;                  /* range [0, nproc-1] or -1                    */
   int partition;
   gid_t adjData[1];           /* TODO: refactor - includes gid & lid & user  */
};

template <typename gid_t, typename lid_t>
class Zoltan2_DD_Find_Msg {  /* Only used by Zoltan_DD_Find()          */
  public:
   int proc;                 /* destination or location                */
   int partition;
   int index;                /* to put things back in order afterward  */
   gid_t adjData[1];         /* TODO: refactor - includes gid and user */
};

template <typename gid_t, typename lid_t>
class Zoltan2_DD_Remove_Msg {  /* Only used by Zoltan_DD_Remove()       */
  public:
   int owner;                  /* range [0, nproc-1] or -1              */
   gid_t adjData[1];           /* TODO: refactor - includes gid         */
};

template <typename gid_t, typename lid_t, typename user_t>
class Zoltan2_Directory {
  public:
    enum Update_Mode {
      Replace = 0, 
      Add, 
      Aggregate, 
      AggregateAdd 
    };

    Zoltan2_Directory(
        Teuchos::RCP<const Teuchos::Comm<int> > comm_,
        bool use_lid_, 
        int debug_level_) 
          : comm(comm_), use_lid(use_lid_),
          debug_level(debug_level_)
    {
    }

    virtual ~Zoltan2_Directory() {
    }

    int update(
      size_t length,                                        /* number of gids */
      const gid_t * gid,                       
      const lid_t * lid,                             
      const user_t * user,                      
      const int * partition,               
      Update_Mode update_mode); 
    int find(
      size_t length,                                        /* number of gids */
      const gid_t * gid,                             
      lid_t * lid,                     
      user_t * user,              
      int * partition,             
      int * owner,                   
      bool throw_if_missing = true); 
    int remove(
      size_t length,                                        /* number of gids */
      const gid_t * gid);                          
    int print() const;

    void stats() const;

    bool is_use_lid() const { return use_lid; }

    void get_locally_managed_gids(std::vector<gid_t> & local_gids) const {
      // resize
      local_gids.resize(node_map.size());

      // fill
      size_t cnt = 0;
      for(size_t n = 0; n < node_map.capacity(); ++n) {
        if(node_map.value_at(n).free == 0) {
          local_gids[cnt++] = node_map.key_at(n);
        }
      }

      if(cnt != node_map.size()) {
        throw std::logic_error("Unexpected counts. Internal error with the"
          " node_map behavior.");
      }
    }

    void remap_user_data_as_unique_gids() {
      // This process follows the pattern of the original unique ids setup
      // It assumes we have updated the directory with keys (as the gid_t) and
      // also created empty user data. Each key is converted to a unique integer
      // and written into the user data.
      typedef long long mpi_t;
      mpi_t nDDEntries = static_cast<mpi_t>(node_map.size());
      mpi_t firstIdx;
      MPI_Scan(&nDDEntries, &firstIdx, 1, MPI_LONG_LONG, MPI_SUM, getRawComm());
      firstIdx -= nDDEntries;  // do not include this rank's entries in prefix sum
      size_t cnt = 0;
      for(size_t n = 0; n < node_map.capacity(); ++n) {
        if(node_map.value_at(n).free == 0) {
          Zoltan2_Directory_Node<gid_t,lid_t,user_t> & node =
            node_map.value_at(n);
          node.userData = firstIdx + cnt;
          cnt++;
        }
      }
    }

    size_t node_map_size() const {
      return node_map.size();
    }

  protected:
    // handled updating the local node information when the proc receives
    // a new gid to store or updated data for a preexisting node
    int update_local(gid_t* gid, lid_t* lid, user_t *user,
      int partition, int owner);

    // collect data on the local proc which has been requested by the directory.
    int find_local(gid_t* gid, lid_t* lid, user_t *user,
      int *partition, int *owner, bool throw_if_missing = true) const;

    // remove the locally stored node for this gid
    int remove_local(gid_t* gid);

    size_t find_msg_size;    /* Total allocation for Zoltan2_DD_FindMsg       */
    size_t update_msg_size;  /* Total allocation for Zoltan2_DD_Update_Msg    */
    size_t remove_msg_size;  /* Total allocation for Zoltan2_DD_Remove_Msg    */

    // originally the nodes are stored in a hash but in the new Kokkos mode
    // they are stored using Kokkos::UnorderedMap
    typedef Kokkos::UnorderedMap<gid_t,
      Zoltan2_Directory_Node<gid_t,lid_t,user_t>, Kokkos::HostSpace> node_map_t;
    node_map_t node_map;

    // this method exists so constructor and copy constructor don't duplicate
    void allocate();

    // this method exists so operator= and copy constructor don't duplicate
    int copy(const Zoltan2_Directory<gid_t,lid_t,user_t> &dd);

    // TODO: Decide if this stays and how to incorporate with variable length
    // data. See comments in Zoltan2_Directory_Impl.hpp
    // size_t align_size_t(size_t a) const;

    // take a gid and hash to proc - determines which proc wil own the gid data
    unsigned int hash_proc(const gid_t & gid) const;

    // stores the comm provided by the user
    Teuchos::RCP<const Teuchos::Comm<int> > comm;

    bool use_lid;                  /* If false not using lid                 */
    int debug_level;               /* Determines actions to multiple updates */

    size_t max_id_size;          /* Stores: max(sizeof(gid_t),sizeof(lid_t)) */
    Update_Mode mode; /* Last mode sent using update */

    // abstract methods are implemented below by Zoltan2_Directory_Simple
    // or Zoltan2_Directory_Vector. These methods contain all the places where
    // the code had to be specialized for normal user type (int, long, etc) or
    // std::vector user type of variable length. Possibly we could consider
    // making this all work by templating but exactly how to do that cleanly
    // I am not sure. The class inheritance approach may be easier to understand
    // but it does mean the user has to pick the right class to use.
    virtual bool is_Zoltan2_Directory_Vector() const                      = 0;
    virtual void update_local_user(const user_t * pRaw, user_t & dst)     = 0;
    virtual void user_to_raw(const user_t & src, user_t * pRaw) const     = 0;
    virtual void raw_to_user(const user_t * pRaw, user_t & dst) const     = 0;
    virtual size_t size_of_value_type() const                             = 0;
    virtual size_t get_update_msg_size(const user_t & data) const         = 0;
    virtual size_t get_update_msg_size(const user_t * pRaw) const         = 0;
    virtual size_t get_local_find_msg_size(gid_t *gid,
      bool throw_if_missing = true) const                                 = 0;
    virtual size_t get_incoming_find_msg_size(
      Zoltan2_DD_Find_Msg<gid_t,lid_t>* msg) const                        = 0;

  private:
    MPI_Comm getRawComm() {
    #ifdef HAVE_MPI
      return Teuchos::getRawMpiComm(*comm);
    #else
      return MPI_COMM_WORLD;
    #endif
    }

    void rehash_node_map(size_t new_hash_size) {
      node_map.rehash(new_hash_size);
    }
};

template <typename gid_t, typename lid_t, typename user_t>
class Zoltan2_Directory_Simple : public Zoltan2_Directory<gid_t, lid_t, user_t> {
  public:
    typedef user_t user_val_t;

    Zoltan2_Directory_Simple(Teuchos::RCP<const Teuchos::Comm<int> > comm_, bool use_lid_,
      int debug_level_) :
      Zoltan2_Directory<gid_t, lid_t, user_t>(comm_, use_lid_,
        debug_level_) {
      // Note that allocate() must be called in the derived class, not the
      // base class or inheritance of the methods will break
      this->allocate();
    }

    Zoltan2_Directory_Simple (
      const Zoltan2_Directory_Simple<gid_t,lid_t,user_t> &src) :
        Zoltan2_Directory<gid_t, lid_t, user_t>(src.comm, src.use_lid,
          src.debug_level) {
      this->allocate();
      this->copy(src);
    }

    Zoltan2_Directory_Simple<gid_t,lid_t,user_t> & operator=
      (const Zoltan2_Directory_Simple<gid_t,lid_t,user_t> &src) {
      this->comm = src.comm;
      this->use_lid = src.use_lid;
      this->debug_level = src.debug_level;
      this->allocate(); // operator= was setup in derived class so this inherits
      this->copy(src);
      return *this;
    }

  protected:
    // awkward to have this at all - so maybe to refactor out with future progress
    virtual bool is_Zoltan2_Directory_Vector() const { return false; }

    // given raw data from the MPI stream we update user data based on mode
    virtual void update_local_user(const user_t * pRaw, user_t & dst) {
      switch(this->mode) {
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::Replace:
          dst = *pRaw;
          break;
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::Add:
           dst += *pRaw;
        break;
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::Aggregate:
          throw std::logic_error("Aggregate doesn't mean anything for single "
            "types. Must use Zoltan2_Directory_Vector class.");
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::AggregateAdd:
          throw std::logic_error("AggregateAdd doesn't mean anything for single "
            "types. Must use Zoltan2_Directory_Vector class.");
      }
    }

    // convert user data to raw data - simple conversion for this class
    virtual void user_to_raw(const user_t & src, user_t * pRaw) const {
      *pRaw = src;
    }

    // convert raw data to user data - simple conversion for this class
    virtual void raw_to_user(const user_t * pRaw, user_t & dst) const {
      dst = *pRaw;
    }

    // get size of the user type which is simply sizeof(user_t) for this class
    virtual size_t size_of_value_type() const { return sizeof(user_t); }

    // for this class, update_msg_size is simple (not variable length)
    virtual size_t get_update_msg_size(const user_t & data) const {
      return this->update_msg_size;
    }

    // for this class, update_msg_size is simple (not variable length)
    virtual size_t get_update_msg_size(const user_t * pRaw) const {
      return this->update_msg_size;
    }

    // for this class, find_msg_size is simple (not variable length)
    virtual size_t get_local_find_msg_size(gid_t *gid,
      bool throw_if_missing = true) const {
      return this->find_msg_size;
    }

    // for this class, find_msg_size is simple (not variable length)
    virtual size_t get_incoming_find_msg_size(
      Zoltan2_DD_Find_Msg<gid_t,lid_t>* msg) const {
      return this->find_msg_size;
    }
};

template <typename gid_t, typename lid_t, typename user_t>
class Zoltan2_Directory_Vector : public Zoltan2_Directory<gid_t, lid_t, user_t> {
  public:
    typedef typename user_t::value_type user_val_t;

    Zoltan2_Directory_Vector(Teuchos::RCP<const Teuchos::Comm<int> > comm_, bool use_lid_,
      int debug_level_) :
      Zoltan2_Directory<gid_t, lid_t, user_t>(comm_, use_lid_,
        debug_level_) {
      // Note that allocate() must be called in the derived class, not the
      // base class or inheritance of the methods will break
      this->allocate();
    }

    Zoltan2_Directory_Vector (
      const Zoltan2_Directory_Vector<gid_t,lid_t,user_t> &src) :
        Zoltan2_Directory<gid_t, lid_t, user_t>(src.comm, src.use_lid,
          src.debug_level) {
      this->allocate(); // operator= was setup in derived class so this inherits
      this->copy(src);
    }

    Zoltan2_Directory_Vector<gid_t,lid_t,user_t> & operator=
      (const Zoltan2_Directory_Vector<gid_t,lid_t,user_t> &src) {
      this->comm = src.comm;
      this->use_lid = src.use_lid;
      this->debug_level = src.debug_level;
      this->allocate(); // operator= was setup in derived class so this inherits
      this->copy(src);
      return *this;
    }

  protected:
    // awkward to have this at all - so maybe to refactor out with future progress
    virtual bool is_Zoltan2_Directory_Vector() const { return true; }

    // given raw data from the MPI stream we update user data based on mode
    virtual void update_local_user(const user_t * pRaw, user_t & dst) {
      // we're reading raw data of form: size_t, val, val, val ...
      size_t * pLength = (size_t*)(pRaw);
      size_t read_array_length = *pLength;
      ++pLength; // move up to first element
      user_val_t * pRead = (user_val_t*)(pLength);
      switch(this->mode) {
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::Replace: {
          // Note this is the raw_to_user method and we could just call it
          // but Add and Aggregate don't have the equivalent so I've done it
          // this way to keep the pattern.
          dst.resize(read_array_length); // change to new
          for(size_t i = 0; i < read_array_length; ++i) {
            dst[i] = *pRead;
            ++pRead;
          }
        }
        break;
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::Add: {
          // ADD currently requires equal length vectors to add each element
          if(dst.size() != static_cast<size_t>(read_array_length)) {
            throw std::logic_error("The data lengths are not the same size");
          }
          // loop through and do the addition
          for(size_t i = 0; i < dst.size(); ++i) {
            dst[i] += *pRead;
            ++pRead;
          }
        }
        break;
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::Aggregate: {
          // Add only unique elements
          // Preserve ordering
          // First scan the new incoming data
          //
          // For example we can have data of:
          //   [1,4,5,7]
          // Then new incoming data is:
          //   [4,5,6,10]
          // Then result would be:
          //   [1,4,5,6,7,10]
          for(size_t i = 0; i < read_array_length; ++i) {
            // handle the cases of dst no size or adding past last element
            if(dst.size() == 0 || (*pRead) > dst[dst.size()-1]) {
              dst.push_back(*pRead); // add first element or at end
            }
            else {
              // otherwise we are going to insert unless it's not unique
              for(auto itr = dst.begin(); itr != dst.end(); ++itr) {
                if((*itr) == (*pRead)) { // do they match
                  if(this->mode == Zoltan2_Directory<gid_t,lid_t,user_t>::
                  Update_Mode::AggregateAdd) {
                    (*itr) += (*pRead); // do the AggregateAdd action using +=
                  }
                  break; // break because it's already in there - do nothing
                }
                else if((*itr) > (*pRead)) { // is scanned element larger?
                  dst.insert(itr, (*pRead)); // preserve ordering
                  break; // break because once we add it we are done
                }
              }
            }
            ++pRead; // get the next incoming array element (*pRead)
          }
        }
        break;
        case Zoltan2_Directory<gid_t,lid_t,user_t>::Update_Mode::AggregateAdd: {
          // AggregateAdd is similar to Aggregate except that when two items
          // match through the operator==, they are combined using operator+=
          // instead of excluding the duplicate. This reliance on operator==
          // and operator+= allows the user to define the behavior for the
          // struct but needs design discussion. This example struct was taken
          // from Zoltan2_GraphMetricsUtility.hpp which was the origina reason
          // for adding this mode.
          /*
              struct part_info {
                part_info() : sum_weights(0) {
                }
                const part_info & operator+=(const part_info & src) {
                  sum_weights += src.sum_weights;
                  return *this;   // return old value
                }
                bool operator>(const part_info & src) {
                  return (target_part > src.target_part);
                }
                bool operator==(const part_info & src) {
                  return (target_part == src.target_part);
                }
                part_t target_part;     // the part this part_info refers to
                t_scalar_t sum_weights; // the sum of weights
              };
          */
          // Then if we use AggregateAdd the following example shows how the
          // struct with part_t 1 will have sum_weights combined:
          //
          //                    Proc 1          Proc 2        Result
          //  part_t           0      1         1    3       0     1     3
          //  sum_weights     1.0    1.0       2.0  2.0     1.0   3.0   2.0
          //
          // TODO: We could make this almost identical to above Aggregate and
          // preserve ordering. Then the only difference is that Aggregate just
          // does nothing when two elements are the same while AggregateAdd will
          // combine them with += operator. Did not implement yet since
          // Zoltan2_GraphMetricsUtility.hpp didn't have parts ordered and I
          // wasn't sure yet if we'd want to make that a requirement.
          for(size_t i = 0; i < read_array_length; ++i) {
            bool bMatch = false;
            for(auto itr = dst.begin(); itr != dst.end(); ++itr) {
              if((*itr) == (*pRead)) { // determine if they go together using ==
                (*itr) += (*pRead); // do the AggregateAdd action using +=
                bMatch = true;
                break;
              }
            }
            if(!bMatch) {
              dst.push_back(*pRead); // add first element or at end
            }
            ++pRead; // get the next incoming array element (*pRead)
          }
        }
        break;

      }
    }

    // write the std::vector as length, x1, x2, x3 ...
    virtual void user_to_raw(const user_t & src, user_t * pRaw) const {
      // we're writing raw data of form: size_t, val, val, val ...
      size_t *pLength = (size_t*)(pRaw);
      *pLength = src.size(); // first write the length
      ++pLength; // move up to first element
      user_val_t *pWrite = (user_val_t*)(pLength);
      for(size_t n = 0; n < src.size(); ++n) {
        *pWrite = src[n]; // now write each element
        ++pWrite;
      }
    }

    // raw comes in as length, x1, x2, x3 ...
    virtual void raw_to_user(const user_t * pRaw, user_t & dst) const {
      // we're reading raw of form: size_t, val, val, val ...
      size_t* pLength = (size_t*) pRaw;
      dst.resize(static_cast<size_t>(*pLength)); // first read the length
      ++pLength; // move up to first element
      user_val_t* pRead = (user_val_t*) pLength;
      for(size_t n = 0; n < dst.size(); ++n) {
        dst[n] = *pRead; // now read each element
        ++pRead;
      }
    }

    // for the std::vector directory, value type is the size of the std::vector
    // template parameter, so for std::vector<int> we want sizeof(int)(
    virtual size_t size_of_value_type() const {
      return sizeof(typename user_t::value_type);
    }

    // the update msg is the base size (includes vector length) plus the size
    // of all the elements.
    virtual size_t get_update_msg_size(const user_t & data) const {
      return this->update_msg_size + data.size() * size_of_value_type();
    }

    // the update msg is the base size (includes vector length) plus the size
    // of all the elements. This is same idea as above method but here we are
    // intepreting raw data (which comes in as length, x1, x2, x3...) so we
    // read the size as the first element. That's all we need to determine the
    // total update_msg_size.
    virtual size_t get_update_msg_size(const user_t * pRaw) const {
      // the first element is size_t (length of the vector)
      size_t * pLength = (size_t*) (pRaw);
      return this->update_msg_size +
        (pRaw ? ((*pLength) * size_of_value_type()) : 0);
    }

    // to get the local find msg size we need to verify the node exists and then
    // read the std::vector size (which is added to base length find_msg_size.
    virtual size_t get_local_find_msg_size(gid_t * gid,
      bool throw_if_missing = true) const {
        if(this->node_map.exists(*gid)) {
          const Zoltan2_Directory_Node<gid_t,lid_t,user_t> & node =
            this->node_map.value_at(this->node_map.find(*gid));
          return this->find_msg_size + node.userData.size() * sizeof(user_val_t);
        }
        else if(throw_if_missing) {
          throw std::logic_error( "Could not find gid in map." );
        }
        else {
          // not clear yet if we ever want to handle this case or always err out
          // I'm using this right now for the unit testing to validate that the
          // remove command actually works.
          return this->find_msg_size; // will not have any data content
        }
    }

    // here we have a find msg coming in and we need to extract the vector length
    // which is always at the same place in the message.
    virtual size_t get_incoming_find_msg_size(
      Zoltan2_DD_Find_Msg<gid_t,lid_t>* msg) const {
      if(msg->proc == -1) {
        // this happens if we called find for an element which was removed
        // eventually we might just throw on find_local but for the testing,
        // this is allowed, the data is left untouched, and the test validates
        // that the results are as expected based on remove events
        return this->find_msg_size; // no extra data for unfound node
      }
      // the first element of the user data is size_t (length of the vector)
      size_t * pVectorLength =
        (size_t*)(reinterpret_cast<char*>(msg->adjData) + this->max_id_size);
      return this->find_msg_size + (*pVectorLength) * sizeof(user_val_t);
    }
};

} // end namespace Zoltan2

#endif