General Memory System
gem5's memory systemwas designed with the following goals:
- Unify timing and functional accesses in timing mode.
- Simplify the memory system code -- remove the huge amount of templating and duplicate code.
- Make changes easier, specifically to allow other memory interconnects besides a shared bus.
- 1 Ports system
- 2 Packets
- 3 Requests
- 4 Atomic/Timing/Functional accesses
- 5 Packet allocation protocol
- 6 Two memory system models: Classic and Ruby
All objects within a memory system inherit from
MemObject. This class adds the pure virtual function
getPort(const std::string &name) which returns a port corresponding to the given name. This interface is used to connect memory objects together with the help of a connector (see below).
The next large part of the memory system is the idea of ports. Ports are used to interface memory objects to each other. They will always come in pairs and we refer to the other port object as the peer. These are used to make the design more modular. With ports a specific interface between every type of object doesn't have to be created. Every memory object has to have at least one port to be useful.
There are two groups of functions in the port object. The
send* functions are called on the port by the object that owns that port. For example to send a packet in the memory system a CPU would call
myPort->sendTiming(pkt) to send a packet. Each send function has a corresponding recv function that is called on the ports peer. So the implementation of the
sendTiming() call above would simply be
peer->recvTiming(pkt). Using this method we only have one virtual function call penalty but keep generic ports that can connect together any memory system objects.
In Python, Ports are first-class attributes of simulation objects, much like Params. Two objects can specify that their ports should be connected using the assignment operator. Unlike a normal variable or parameter assignment, port connections are symmetric:
A.port1 = B.port2 has the same meaning as
B.port2 = A.port1.
Objects such as busses that have a potentially unlimited number of ports use "vector ports". An assignment to a vector port appends the peer to a list of connections rather than overwriting a previous connection.
In C++, memory ports are connected together by the python code doing something like
p1 = obj1->getPort(); p2 = obj2->getPort(); p1->setPeer(p2); p2->setPeer(p1); with the appropriate port names if any. This is done after all objects are instantiated.
There are several types of ports that inherit from
Port and are very generic.
FunctionalPortprovides easy to use methods for writing and reading physical addresses. It is only meant to load data into memory and update constants before the simulation begins.
VirtualPortprovides the same methods as
FunctionalPort, but the addresses passed to it are virtual addresses, and a translation is done to get the physical address. If no
ThreadContextis passed to the constructor, the virtual->physical translation must be static (e.g. Alpha Superpage accesses), otherwise a
ThreadContextis required to do the translation.
A Packet is used to encapsulate a transfer between two objects in the memory system (e.g., the L1 and L2 cache). This is in contrast to a Request where a single Request travels all the way from the requester to the ultimate destination and back, possibly being conveyed by several different Packets along the way.
Read access to many packet fields is provided via accessor methods which verify that the data in the field being read is valid.
A packet contains the following all of which are accessed by accessors to be certain the data is valid:
- The address. This is the address that will be used to route the packet to its target (if the destination is not explicitly set) and to process the packet at the target. It is typically derived from the request object's physical address, but may be derived from the virtual address in some situations (e.g., for accessing a fully virtual cache before address translation has been performed). It may not be identical to the original request address: for example, on a cache miss, the packet address may be the address of the block to fetch and not the request address.
- The size. Again, this size may not be the same as that of the original request, as in the cache miss scenario.
- A pointer to the data being manipulated.
- Set by
dataDynamicArray()which control if the data associated with the packet is freed when the packet is, not, with
delete, and with
- Allocated if not set by one of the above methods
allocate()and the data is freed when the packet is destroyed. (Always safe to call).
- A pointer can be retrived by calling
set()can be used to manipulate the data in the packet. The get() method does a guest-to-host endian conversion and the set method does a host-to-guest endian conversion.
- Set by
- A status indicating Success, BadAddress, Not Acknowleged, and Unknown.
- A list of command attributes associated with the packet
- Note: There is some overlap in the data in the status field and the command attributes. This is largely so that a packet an be easily reinitialized when nacked or easily reused with atomic or functional accesses.
SenderStatepointer which is a virtual base opaque structure used to hold state associated with the packet but specific to the sending device (e.g., an MSHR). A pointer to this state is returned in the packet's response so that the sender can quickly look up the state needed to process it. A specific subclass would be derived from this to carry state specific to a particular sending device.
CoherenceStatepointer which is a virtual base opaque structure used to hold coherence-related state. A specific subclass would be derived from this to carry state specific to a particular coherence protocol.
- A pointer to the request.
A request object encapsulates the original request issued by a CPU or I/O device. The parameters of this request are persistent throughout the transaction, so a request object's fields are intended to be written at most once for a given request. There are a handful of constructors and update methods that allow subsets of the object's fields to be written at different times (or not at all). Read access to all request fields is provided via accessor methods which verify that the data in the field being read is valid.
The fields in the request object are typically not available to devices in a real system, so they should normally be used only for statistics or debugging and not as architectural values.
Request object fields include:
- Virtual address. This field may be invalid if the request was issued directly on a physical address (e.g., by a DMA I/O device).
- Physical address.
- Data size.
- Time the request was created.
- The ID of the CPU/thread that caused this request. May be invalid if the request was not issued by a CPU (e.g., a device access or a cache writeback).
- The PC that caused this request. Also may be invalid if the request was not issued by a CPU.
There are three types of accesses supported by the ports.
- Timing - Timing accesses are the most detailed access. They reflect our best effort for realistic timing and include the modeling of queuing delay and resource contention. Once a timing request is successfully sent at some point in the future the device that sent the request will either get the response or a NACK if the request could not be completed (more below). Timing and Atomic accesses can not coexist in the memory system.
- Atomic - Atomic accesses are a faster than detailed access. They are used for fast forwarding and warming up caches and return an approximate time to complete the request without any resource contention or queuing delay. When a atomic access is sent the response is provided when the function returns. Atomic and timing accesses can not coexist in the memory system.
- Functional - Like atomic accesses functional accesses happen instantaneously, but unlike atomic accesses they can coexist in the memory system with atomic or timing accesses. Functional accesses are used for things such as loading binaries, examining/changing variables in the simulated system, and allowing a remote debugger to be attached to the simulator. The important note is when a functional access is received by a device, if it contains a queue of packets all the packets must be searched for requests or responses that the functional access is effecting and they must be updated as appropriate. The
fixPacket()methods can help with this.
Timing Flow control
Timing requests simulate a real memory system, so unlike functional and atomic accesses their response is not instantaneous. Because the timing requests are not instantaneous, flow control is needed. When a timing packet is sent via
sendTiming() the packet may or may not be accepted, which is signaled by returning true or false. If false is returned the object should not attempt to sent anymore packets until it receives a
recvRetry() call. At this time it should again try to call
sendTiming(); however the packet may again be rejected. Note: The original packet does not need to be resent, a higher priority packet can be sent instead. Once
sendTiming() returns true, the packet may still not be able to make it to its destination. For packets that require a response (i.e.
pkt->needsResponse() is true), any memory object can refuse to acknowledge the packet by changing its result to
Nacked and sending it back to its source. However, if it is a response packet, this can not be done. The true/false return is intended to be used for local flow control, while nacking is for global flow control. In both cases a response can not be nacked.
Response and Snoop ranges
Ranges in the memory system are handled by having devices that are sensitive to an address range provide an implementation for
getDeviceAddressRanges in their port objects. This method returns two
AddrRangeLists: one of addresses it responds to and one of addresses it snoops. When these ranges change (e.g. from PCI configuration taking place) the device should call
sendStatusChange(Port::RangeChange) on its port so that the new ranges are propagated to the entire hierarchy. This is precisely what happens during
init(); all memory objects call
sendStatusChange(Port::RangeChange), and a flurry of range updates occur until everyones ranges have been propagated to all busses in the system.
Packet allocation protocol
The protocol for allocation and deallocation of Packet objects varies depending on the access type. (We're talking about low-level C++
delete issues here, not anything related to the coherence protocol.)
- Atomic and Functional
- The Packet object is owned by the requester. The responder must overwrite the request packet with the response (typically using the
Packet::makeResponse()method). There is no provision for having multiple responders to a single request. Since the response is always generated before
sendFunctional()returns, the requester can allocate the Packet object statically or on the stack.
- Timing transactions are composed of two one-way messages, a request and a response. In both cases, the Packet object must be dynamically allocated by the sender. Deallocation is the responsibility of the receiver (or, for broadcast coherence packets, the target device, typically memory). In the case where the receiver of a request is generating a response, it may choose to reuse the request packet for its response to save the overhead of calling
new(and gain the convenience of using
makeResponse()). However, this optimization is optional, and the requester must not rely on receiving the same Packet object back in response to a request. Note that when the responder is not the target device (as in a cache-to-cache transfer), then the target device will still delete the request packet, and thus the responding cache must allocate a new Packet object for its response. Also, because the target device may delete the request packet immediately on delivery, any other memory device wishing to reference a broadcast packet past point where the packet is delivered must make a copy of that packet, as the pointer to the packet that is delivered cannot be relied upon to stay valid.
Two memory system models: Classic and Ruby
The gem5 simulator includes two different memory system models, Classic and Ruby, that incorporate the above mentioned general memory system components. As the name suggests, the Classic memory system model is inherited from the previous M5 simulator, while the Ruby memory system model is based on the GEMS memory system model of the same name. Each model is complementary and both have unique advantages and disadvantages. This sub-section introduces each model and compares their functionality. Then the next two respective top-level sections describe each model in detail.
Classic memory system
The Classic memory system model provides gem5 a fast and easily configurable memory system at the cost of accuracy and flexibility. All objects within the Classic model inherit from
MemObject and connect together using ports. In particular, ports support direct point-to-point connections between two
MemObjects and buses connect two or more
MemObjects together. Cache coherence is maintained using an abstract snooping protocol where probes across the memory hierarchy functionally occur instantaneously. Using this methodology, the Classic memory system model has the following advantages and disadvantages:
- Fast Forwarding - The Classic model supports atomic accesses, which as stated above, are faster than detailed accesses. This mode of operation is especially advantageous when one needs to fast forward to interesting parts of the execution.
- Speed - Not only does the Classic model support fast atomic accesses, but its timing accesses are relatively fast as compared to Ruby.
- Ease of Configuration - By simply modifying the python configuration, the Classic model allows one to create an arbitrary memory hierarchy. The Classic's abstract cache coherence protocol automatically extends to any memory hierarchy as long as it is composed of caches, buses, and cpus.
- Cache Coherence Flexibility - While the Classic model allows one to create arbitrary systems composed of caches, buses, and cpus, the Classic model is restricted to its abstract snooping protocol. Modifying the protocol requires significant effort.
- Accuracy - By not modeling transient states and the timing of probes, the Classic model does not model protocol contention as accurately as the Ruby model.
Ruby memory system
In contrast to the Classic model, the Ruby memory system model sacrifices simulation speed to provide gem5 a flexible infrastructure capable of accurately simulating a wide variety of memory systems. In particular, Ruby supports a domain specific language called SLICC (Specification Language for Implementing Cache Coherence) where one can define many different types of cache coherence protocols. Essentially SLICC defines the cache, memory, and dma controllers as individual per-memory-block state machines that together form the overall protocol. By defining the controller logic in a higher level language, SLICC allows different protocols to incorporate the same underlining state transition mechanisms with minimal programmer effort.
Unlike the Classic model, Ruby does not connect all objects together using ports. Instead, ports only connect cpus and devices to the memory system via the
RubyPort object. Then within the Ruby memory system, all objects are connected to each other via
MessageBuffers. MessageBuffers are similar to ports in that they provide objects a standard communication interface. However, MessageBuffers include a queue that stores messages while ports do not. Messages cannot be enqueued and dequeued at the same simulated cycle and thus communication across a message buffer is not instantaneous. The result is MessageBuffers only support timing accesses and cannot support the instantaneous atomic and functional accesses.
To summarize, the Ruby memory system model has the following advantages and disadvantages:
- Cache Coherence Flexibility - Utilizing SLICC, Ruby can implement a wide variety of cache coherence protocols, from directory to snooping protocols and several points in between.
- Accuracy - Ruby accurately models both cache coherence and network related features in the memory system. In particular, SLICC's message trigger event methodology accurately models transient state timing. Also the Garnet network model integrated in Ruby accurately models network contention and flow control.
- Fast Forwarding - Ruby does not support atomic accesses and thus does not have reasonable fast forwarding capability.
- Speed - As compared to the Classic model, Ruby is relatively slow. This is especially true when using the Garnet network model.
- Ease of Configuration - While SLICC is a powerful tool to model a wide variety of cache coherence protocols, the resulting protocols are optimized for a specific cache hierarchy configuration. As such, it is difficult to simply extend protocols to another level of cache.