Server Implementation Techniques

As we mentioned on page 856, instantiating a servant for each Ice object on server start‑up is a viable design, provided that you can afford the amount of memory required by the servants, as well as the delay in start‑up of the server. However, Ice supports more flexible mappings between Ice objects and servants; these alternate mappings allow you to precisely control the trade-off between memory consumption, scalability, and performance. We outline a few of the more common implementation techniques in this section.

32.9.1 Incremental Initialization

If you use a servant locator, the servant returned by locate is used only for the current request, that is, the Ice run time does not add the servant to the active servant map. Of course, this means that if another request comes in for the same Ice object, locate must again retrieve the object state and instantiate a servant. A common implementation technique is to add each servant to the ASM as part of locate. This means that only the first request for each Ice object triggers a call to locate; thereafter, the servant for the corresponding Ice object can be found in the ASM and the Ice run time can immediately dispatch another incoming request for the same Ice object without having to call the servant locator.

Ice::ObjectPtr
MyServantLocator::locate(const Ice::Current&  c,
                         Ice::LocalObjectPtr& cookie)
{
    // Get the object identity. (We use the name member
    // as the database key.)
    //
    std::string name = c.id.name;

    // Use the identity to retrieve the state from the database.
    //
    ServantDetails d;
    try {
        d = DB_lookup(name);
    } catch (const DB_error&)
        return 0;
    }

    // We have the state, instantiate a servant.
    //
    Ice::ObjectPtr servant = new PhoneEntryI(d);

    // Add the servant to the ASM.
    //
    c.adapter‑>add(servant, c.id);      // NOTE: Incorrect!

    return servant;
}

This is almost identical to the implementation on page 864—the only difference is that we also add the servant to the ASM by calling ObjectAdapter::add. Unfortunately, this implementation is wrong because it suffers from a race condition. Consider the situation where we do not have a servant for a particular Ice object in the ASM, and two clients more or less simultaneously send a request for the same Ice object. It is entirely possible for the thread scheduler to schedule the two incoming requests such that the Ice run time completes the lookup in the ASM for both requests and, for each request, concludes that no servant is in memory. The net effect is that locate will be called twice for the same Ice object, and our servant locator will instantiate two servants instead of a single servant. Because the second call to ObjectAdapter::add will raise an AlreadyRegisteredException, only one of the two servants will be added to the ASM.

Of course, this is hardly the behavior we expect. To avoid the race condition, our implementation of locate must check whether a concurrent invocation has already instantiated a servant for the incoming request and, if so, return that servant instead of instantiating a new one. The Ice run time provides the ObjectAdapter::find operation to allow us to test whether an entry for a specific identity already exists in the ASM:

find returns the servant if it exists in the ASM and null, otherwise. Using this lookup function, together with a mutex, allows us to correctly implement locate. The class definition of our servant locator now has a private mutex so we can establish a critical region inside locate:

class MyServantLocator : public virtual Ice::ServantLocator {
public:

    virtual Ice::ObjectPtr locate(const Ice::Current&  c,
                                  Ice::LocalObjectPtr&);

    // Declaration of finished() and deactivate() here...

private:
    IceUtil::Mutex _m;
};

Ice::ObjectPtr
MyServantLocator::locate(const Ice::Current&  c,
                         Ice::LocalObjectPtr&)
{
    IceUtil::Mutex::Lock lock(_m);

    // Check if we have instantiated a servant already.
    //
    Ice::ObjectPtr servant = c.adapter.find(c.id);

    if (!servant) {     // We don't have a servant already

        // Instantiate a servant.
        //
        ServantDetails d;
        try {
           d = DB_lookup(c.id.name);
        } catch (const DB_error&) {
           return 0;
        }
        servant = new PhoneEntryI(d);

        // Add the servant to the ASM.
        //
        c.adapter‑>add(servant, c.id);
    }

    return servant;
}

synchronized public Ice.Object
locate(Ice.Current c, Ice.LocalObjectHolder cookie)
{
    // Check if we have instantiated a servant already.
    //
    Ice.Object servant = c.adapter.find(c.id);

    if (servant == null) { // We don't have a servant already

        // Instantiate a servant
        //
        ServantDetails d;
        try {
            d = DB.lookup(c.id.name);
        } catch (DB.error&) {
            return null;
        }
        servant = new PhoneEntryI(d);

        // Add the servant to the ASM.
        //
        c.adapter.add(servant, c.id);
    }

    return servant;
}

• The memory requirements for the server are reduced because servants are instantiated only for those Ice objects that are actually accessed by clients. If clients only access a subset of the total number of Ice objects, the memory savings can be substantial.

In general, incremental initialization is beneficial if instantiating servants during start‑up is too slow. The memory savings can be worthwhile as well but, as a rule, are realized only for comparatively short-lived servers: for long-running servers, chances are that, sooner or later, every Ice object will be accessed by some client or another; in that case, there are no memory savings because we end up with an instantiated servant for every Ice object regardless.

32.9.2 Default Servants

As explained in Section 32.8, default servants are a very effective tool for conserving memory when a server hosts a large number of Ice objects. Some Ice products support a dedicated API for default servants, but the same functionality can be implemented using servant locators with just a little more effort.

To create a default servant implementation with servant locators, we need as many locators as there are non-abstract interfaces in the system. For example, for our file system application, we require two locators, one for directories and one for files. In addition, the object identities we create use the category member of the object identity to encode the type of interface of the corresponding Ice object. The value of the category field can be anything that identifies the interface, such as the ‘d’ and ‘f’ convention we used on page 868. Alternatively, you could use "Directory" and "File", or use the type ID of the corresponding interface, such as "::Filesystem::Directory" and "::Filesystem::File". The name member of the object identity must be set to whatever identifier we can use to retrieve the persistent state of each directory and file from secondary storage. (For our file system application, we used a UUID as a unique identifier.)

class DirectoryLocator : public virtual Ice::ServantLocator {
public:
    DirectoryLocator() : _servant(new DirectoryI)
    {
    }

    virtual Ice::ObjectPtr locate(const Ice::Current&,
                                  Ice::LocalObjectPtr&)
    {
        return _servant;
    }

    virtual void finished(const Ice::Current&,
                          const Ice::ObjectPtr&,
                          const Ice::LocalObjectPtr&)
    {
    }

    virtual void deactivate(const std::string&)
    {
    }

private:
    Ice::ObjectPtr _servant;
};

Note that constructor of the locator instantiates a servant and returns that same servant from every call to locate. The implementation of the file locator is analogous to the one for directories. Registration of the servant locators proceeds as usual:

_adapter‑>addServantLocator(new DirectoryLocator, "d");
_adapter‑>addServantLocator(new FileLocator, "f");

Filesystem::NodeSeq
Filesystem::DirectoryI::list(const Ice::Current& c) const
{
    // Use the identity of the directory to retrieve
    // its contents.
    DirectoryContents dc;
    try {
        dc = DB_getDirectory(c.id.name);
    } catch(const DB_error&) {
        throw Ice::ObjectNotExistException(__FILE__, __LINE__);
    }

    // Use the records retrieved from the database to
    // initialize return value.
    //
    FileSystem::NodeSeq ns;
    // ...

    return ns;
}

Overriding ice_ping

Section 32.8.5 recommends that a default servant implementation take steps to preserve the semantics of the ice_ping operation, which is used to test whether an Ice object exists. If a default servant fails to override ice_ping, clients may mistakenly believe that a non-existent Ice object still exists. The code below demonstrates how we can override the operation in our file system application:

void
Filesystem::DirectoryI::ice_ping(const Ice::Current& c) const
{
    try {
       d = DB_lookup(c.id.name);
    } catch (const DB_error&) {
       throw Ice::ObjectNotExistException(__FILE__, __LINE__);
    }
}

32.9.3 Hybrid Approaches and Caching

Depending on the nature of your application, you may be able to steer a middle path that provides better performance while keeping memory requirements low: if your application has a number of frequently-accessed objects that are performance-critical, you can add servants for those objects to the ASM. If you store the state of these objects in data members inside the servants, you effectively have a cache of these objects.

The remaining, less-frequently accessed objects can be implemented with a default servant. For example, in our file system implementation, we could choose to instantiate directory servants permanently, but to have file objects implemented with a default servant. This provides efficient navigation through the directory tree and incurs slower performance only for the (presumably less frequent) file accesses.

This technique could be augmented with a cache of recently-accessed files, along similar lines to the buffer pool used by the Unix kernel [10]. The point is that you can combine use of the ASM with servant locators and default servants to precisely control the trade-offs among scalability, memory consumption, and performance to suit the needs of your application.

32.9.4 Servant Evictors

• Whenever a request arrives (that is, locate is called by the Ice run time), the evictor checks to see whether it can find a servant for the request in its cache. If so, it returns the servant that is already instantiated in the cache; otherwise, it instantiates a servant and adds it to the cache.

• The cache is a queue that is maintained in least-recently used (LRU) order: the least-recently used servant is at the tail of the queue, and the most-recently used servant is at the head of the queue. Whenever a servant is returned from or added to the cache, it is moved from its current queue position to the head of the queue, that is, the “newest” servant is always at the head, and the “oldest” servant is always at the tail.

• The queue has a configurable length that corresponds to how many servants will be held in the cache; if a request arrives for an Ice object that does not have a servant in memory and the cache is full, the evictor removes the least-recently used servant at the tail of the queue from the cache in order to make room for the servant about to be instantiated at the head of the queue.

At this point, the evictor has instantiated five servants, and has placed each servant onto the evictor queue. Because requests were sent by the client for object identities 1 to 5 (in that order), servant 5 ends up at the head of the queue (at the most-recently used position), and servant 1 ends up at the tail of the queue (at the least-recently used position).

Assume that the next client request is for object identity 6. The evictor queue is fully populated, so the evictor creates a servant for object identity 6, places that servant at the head of the queue, and evicts the servant with identity 1 (the least-recently used servant) at the tail of the queue, as shown in Figure 32.4.

The evictor pattern combines the advantages of the ASM with the advantages of a default servant: provided that the cache size is sufficient to hold the working set of servants in memory, most requests are served by an already instantiated servant, without incurring the overhead of creating a servant and accessing the database to initialize servant state. By setting the cache size, you can control the trade-off between performance and memory consumption as appropriate for your application.

Creating an Evictor Implementation in C++

The evictor we show here is designed as an abstract base class: in order to use it, you derive an object from the EvictorBase base class and implement two methods that are called by the evictor when it needs to add or evict a servant. This leads to a class definitions as follows:

class EvictorBase : public Ice::ServantLocator {
public:
    EvictorBase(int size = 1000);

    virtual Ice::ObjectPtr locate(const Ice::Current& c,
                                  Ice::LocalObjectPtr& cookie);
    virtual void finished(const Ice::Current& c,
                          const Ice::ObjectPtr&,
                          const Ice::LocalObjectPtr& cookie);
    virtual void deactivate(const std::string&);

protected:
    virtual Ice::ObjectPtr add(const Ice::Current&,
                               Ice::LocalObjectPtr&) = 0;
    virtual void evict(const Ice::ObjectPtr&,
                       const Ice::LocalObjectPtr&) = 0;

private:
    // ...
};

typedef IceUtil::Handle<EvictorBase> EvictorBasePtr;

The add and evict functions are called by the evictor when it needs to add a new servant to the queue and when it evicts a servant from the queue. Note that these functions are pure virtual, so they must be implemented in a derived class. The job of add is to instantiate and initialize a servant for use by the evictor. The evict function is called by the evictor when it evicts a servant. This allows evict to perform any cleanup. Note that add can return a cookie that the evictor passes to evict, so you can move context information from add to evict.

2. The map stores an iterator into the evictor queue that marks the position of the servant in the queue. Storing the queue position is not strictly necessary—we store the position for efficiency reasons because it allows us to locate a servant’s position in the queue in constant time instead of having to search through the queue in order to maintain its LRU property.

The need for the use count deserves some extra explanation: suppose a client invokes a long-running operation on an Ice object with identity I. In response, the evictor adds a servant for I to the evictor queue. While the original invocation is still executing, other clients invoke operations on various Ice objects, which leads to more servants for other object identities being added to the queue. As a result, the servant for identity I gradually migrates toward the tail of the queue. If enough client requests for other Ice objects arrive while the operation on object I is still executing, the servant for I could be evicted while it is still executing the original request.

By itself, this will not do any harm. However, if the servant is evicted and a client then invokes another request on object I, the evictor would have no idea that a servant for I is still around and would add a second servant for I. However, having two servants for the same Ice object in memory is likely to cause problems, especially if the servant’s operation implementations write to a database.

The use count allows us to avoid this problem: we keep track of how many requests are currently executing inside each servant and, while a servant is busy, avoid evicting that servant. As a result, the queue size is not a hard upper limit: long-running operations can temporarily cause more servants than the limit to appear in the queue. However, as soon as excess servants become idle, they are evicted as usual.

The evictor queue does not store the identity of the servant. Instead, the entries on the queue are iterators into the evictor map. This is useful when the time comes to evict a servant: instead of having to search the map for the identity of the servant to be evicted, we can simply delete the map entry that is pointed at by the iterator at the tail of the queue. We can get away with storing an iterator into the evictor queue as part of the map, and storing an iterator into the evictor map as part of the queue because both std::list and std::map do not invalidate forward iterators when we add or delete entries² (except for invalidating iterators that point at a deleted entry, of course).

class EvictorBase : public Ice::ServantLocator {
    // ...

private:

    struct EvictorEntry;
    typedef IceUtil::Handle<EvictorEntry> EvictorEntryPtr;

    typedef std::map<Ice::Identity, EvictorEntryPtr> EvictorMap;
    typedef std::list<EvictorMap::iterator> EvictorQueue;

    struct EvictorEntry : public Ice::LocalObject
    {
        Ice::ObjectPtr servant;
        Ice::LocalObjectPtr userCookie;
        EvictorQueue::iterator queuePos;
        int useCount;
    };

    EvictorMap _map;
    EvictorQueue _queue;
    Ice::Int _size;

    IceUtil::Mutex _mutex;

    void evictServants();
};

Note that the evictor stores the evictor map, queue, and the queue size in the private data members _map, _queue, and _size. In addition, we use a private _mutex data member so we can correctly serialize access to the evictor’s data structures.

EvictorBase::EvictorBase(Ice::Int size) : _size(size)
{
if (_size < 0)
_size = 1000;
}

Ice::ObjectPtr
EvictorBase::locate(const Ice::Current& c,
                    Ice::LocalObjectPtr& cookie)
{
    IceUtil::Mutex::Lock lock(_mutex);

    //
    // Check if we have a servant in the map already.
    //
    EvictorEntryPtr entry;
    EvictorMap::iterator i = _map.find(c.id);
    if (i != _map.end()) {
        //
        // Got an entry already, dequeue the entry from
        // its current position.
        //
        entry = i‑>second;
        _queue.erase(entry‑>queuePos);
    } else {
        //
        // We do not have an entry. Ask the derived class to
        // instantiate a servant and add a new entry to the map.
        //
        entry = new EvictorEntry;
        entry‑>servant = add(c, entry‑>userCookie); // Down‑call
        if (!entry‑>servant) {
            return 0;
        }
        entry‑>useCount = 0;
        i = _map.insert(std::make_pair(c.id, entry)).first;
    }

    //
    // Increment the use count of the servant and enqueue
    // the entry at the front, so we get LRU order.
    //
    ++(entry‑>useCount);
    entry‑>queuePos = _queue.insert(_queue.begin(), i);

    cookie = entry;

    return entry‑>servant;
}

The first step in locate is to lock the _mutex data member. This protects the evictor’s data structures from concurrent access. The next step is to instantiate a smart pointer to an EvictorEntry. That smart pointer acts as the cookie that is returned from locate and will be passed by the Ice run time to the corresponding call to finished. That same smart pointer is also the value type of our map entries, so we do not store two copies of the same information redundantly—instead, smart pointers ensure that a single copy of each EvictorEntry structure is shared by both the cookie and the map.

Otherwise, we do not have an entry for this object identity yet, so we have to create one. The code creates a new evictor entry, and then calls add to get a new servant. This is a down-call to the concrete class that will be derived from EvictorBase. The implementation of add must attempt to locate the object state for the Ice object with the identity passed inside the Current object and either return a servant as usual, or return null or throw an exception to indicate failure. If add returns null, we return zero to let the Ice run time know that no servant could be found for the current request. If add succeeds, we initialize the entry’s use count to zero and insert the entry into the evictor map.

The last few lines of locate add the entry for the current request to the head of the evictor queue to maintain its LRU property, increment the use count of the entry, set the cookie that is returned from locate to point at the EvictorEntry, and finally return the servant to the Ice run time.

void
EvictorBase::finished(const Ice::Current&,
                      const Ice::ObjectPtr&,
                      const Ice::LocalObjectPtr& cookie)
{
    IceUtil::Mutex::Lock lock(_mutex);

    EvictorCookiePtr ec = EvictorCookiePtr::dynamicCast(cookie);

    // Decrement use count and check if
    // there is something to evict.
    //
    ‑‑(ec‑>entry‑>useCount);
    evictServants();
}

void
EvictorBase::evictServants()
{
    //
    // If the evictor queue has grown larger than the limit,
    // look at the excess elements to see whether any of them
    // can be evicted.
    //
    EvictorQueue::reverse_iterator p = _queue.rbegin();
    int excessEntries = static_cast<int>(_map.size() ‑ _size);

    for (int i = 0; i < excessEntries; ++i) {
        EvictorMap::iterator mapPos = *p;
        if (mapPos‑>second‑>useCount == 0) {
            evict(mapPos‑>second‑>servant,
                  mapPos‑>second‑>userCookie);
            p = EvictorQueue::reverse_iterator(
                    _queue.erase(mapPos‑>second‑>queuePos));
            _map.erase(mapPos);
        } else
            ++p;
    }
}

The code scans the excess entries, starting at the tail of the evictor queue. If an entry has a zero use count, it is evicted: after calling the evict member function in the derived class, the code removes the evicted entry from both the map and the queue.

Finally, the implementation of deactivate sets the evictor size to zero and then calls evictServants. This results in eviction of all servants. The Ice run time guarantees to call deactivate only once no more requests are executing in an object adapter; as a result, it is guaranteed that all entries in the evictor will be idle and therefore will be evicted.

void
EvictorBase::deactivate(const std::string& category)
{
    IceUtil::Mutex::Lock lock(_mutex);

    _size = 0;
    evictServants();
}