Patent Application
20080186998
The invention relates to an integrated circuit having a plurality of processing modules and a network arranged for coupling processing modules and a method for time slot allocation in such an integrated circuit, and a data processing system.
Systems on silicon show a continuous increase in complexity due to the ever increasing need for implementing new features and improvements of existing functions. This is enabled by the increasing density with which components can be integrated on an integrated circuit. At the same time the clock speed at which circuits are operated tends to increase too. The higher clock speed in combination with the increased density of components has reduced the area which can operate synchronously within the same clock domain. This has created the need for a modular approach. According to such an approach a processing system comprises a plurality of relatively independent, complex modules. In conventional processing systems the modules usually communicate to each other via a bus. As the number of modules increases however, this way of communication is no longer practical for the following reasons. A large number of modules represent a high bus load. Further the bus represents a communication bottleneck as it enables only one module to send data to the bus.
A communication network forms an effective way to overcome these disadvantages.
Networks on chip (NoC) have received considerable attention recently as a solution to the interconnection problem in highly-complex chips. The reason is twofold. First, NoCs help resolve the electrical problems in new deep-submicron technologies, as they structure and manage global wires. At the same time the NoC concept share wires, allows a reduction of the number of wires and increases the utilization of wires. NoCs can also be energy efficient and reliable and are scalable compared to buses. Second, NoCs also decouple computation from communication, which is essential in managing the design of billion-transistor chips. NoCs achieve this decoupling because they are traditionally designed using protocol stacks, which provide well- defined interfaces separating communication service usage from service implementation.
Introducing networks as on-chip interconnects radically changes the communication when compared to direct interconnects, such as buses or switches. This is because of the multi-hop nature of a network, where communication modules are not directly connected, but are remotely separated by one or more network nodes. This is in contrast with the prevalent existing interconnects (i.e., buses) where modules are directly connected. The implications of this change reside in the arbitration (which must change from centralized to distributed), and in the communication properties (e.g., ordering, or flow control), which must be handled either by an intellectual property block (IP) or by the network.
Most of these topics have been already the subject of research in the field of local and wide area networks (computer networks) and as an interconnect for parallel processor networks. Both are very much related to on-chip networks, and many of the results in those fields are also applicable on chip. However, NoC's premises are different from off-chip networks, and, therefore, most of the network design choices must be reevaluated. On-chip networks have different properties (e.g., tighter link synchronization) and resource constraints (e.g., higher memory cost) leading to different design choices, which ultimately affect the network services. Storage (i.e., memory) and computation resources are relatively more expensive, whereas the number of point-to-point links is larger on chip than off chip. Storage is expensive, because general-purpose on-chip memory, such as RAMs, occupies a large area. Having the memory distributed in the network components in relatively small sizes is even worse, as the overhead area in the memory then becomes dominant.
A network on chip (NoC) typically consists of a plurality of routers and network interfaces. Routers serve as network nodes and are used to transport data from a source network interface to a destination network interface by routing data on a correct path to the destination on a static basis (i.e., route is predetermined and does not change), or on a dynamic basis (i.e., route can change depending e.g., on the NoC load to avoid hot spots). Routers can also implement time guarantees (e.g., rate-based, deadline-based, or using pipelined circuits in a TDMA fashion). A known example for NoCs is AEthereal.
The network interfaces are connected to processing modules, also called IP blocks, which may represent any kind of data processing unit, a memory, a bridge, a compressor etc. In particular, the network interfaces constitute a communication interface between the processing modules and the network. The interface is usually compatible with the existing bus interfaces. Accordingly, the network interfaces are designed to handle data sequentialization (fitting the offered command, flags, address, and data on a fixed-width (e.g., 32 bits) signal group) and packetization (adding the packet headers and trailers needed internally by the network). The network interfaces may also implement packet scheduling, which may include timing guarantees and admission control.
An NoC provides various services to processing modules to transfer data between them.
The NoC could be operated according to best effort (BE) or guaranteed throughput (GT) services. In best effort (BE) service, there are no guarantees about latency or throughput. Data is forwarded through routers without any reservation of slots. So this kind of data faces contention in the router, whereas giving guarantees is not possible. In contrast, GT service allows deriving exact value for latency and throughput for transmitting data between processing modules.
On-chip systems often require timing guarantees for their interconnect communications. A cost-effective way of providing time-related guarantees (i.e., throughput, latency and jitter) is to use pipelined circuits in a TDMA (Time Division Multiple Access) fashion, which is advantageous as it requires less buffer space compared to rate-based and deadline-based schemes on systems on chip (SoC) which have tight synchronization. Therefore, a class of communication is provided, in which throughput, latency and jitter are guaranteed, based on a notion of global time (i.e., a notion of synchronicity between network components, i.e. routers and network interfaces), wherein the basic time unit is called a slot or time slot. All network components usually comprise a slot table of equal size for each output port of the network component, in which time slots are reserved for different connections.
At the transport layer of the network, the communication between the processing modules is performed over connections. A connection is considered as a set of channels, each having a set of connection properties, between a first processing module and at least one second processing module. For a connection between a first processing module and a single second processing module, the connection may comprises two channels, namely one from the first to the second processing module, i.e. the request or forward channel, and a second channel from the second to the first processing module, i.e. the response or reverse channel. The forward or request channel is reserved for data and messages from the master to the slave, while the reverse or response channel is reserved for data and messages from the slave to the master. If no response is required, the connection may only comprise one channel. It is not illustrated but possible, that the connection involves one master and N slaves. In that case 2*N channels are provided. Therefore, a connection or the path of the connection through the network comprises at least one channel. In other words, a channel corresponds to the connection path of the connection if only one channel is used. If two channels are used as mentioned above, one channel will provide the connection path e.g. from the master to the slave, while the second channel will provide the connection path from the slave to the master. Accordingly, for a typical connection, the connection path will comprise two channels. The connection properties may include ordering (data transport in order), flow control (a remote buffer is reserved for a connection, and a data producer will be allowed to send data only when it is guaranteed that buffer space is available for the produced data), throughput (a lower bound on throughput is guaranteed), latency (upper bound for latency is guaranteed), the lossiness (dropping of data), transmission termination, transaction completion, data correctness, priority, or data delivery. In NoC, connections are built on top channels. A channel is a unidirectional path through the network from a source (master, initiator) to a destination (slave, target) or back.
For implementing GT services slot tables are used. The slot tables as mentioned above are stored in the network components, including network interfaces and routers. The slot tables allow a sharing of the same link or wires in a time-division multiple access, TDMA, manner. The quantum of data that is injected into the network is called a flit, wherein a flit is a fixed size sub-packet. The injection of flits is regulated by the slot table stored in the network interface. The slot table advances in synchronization (i.e., all are in the same slot at the same time). A channel may have one or more slots allocated within a slot table. The slot tables in all network components are so filled that flits communicated over the network do not content. The channels are used to identify different traffic classes and associate properties to them. At each slot, a data item is moved from one network component to the next one, i.e. between routers or between a router and a network interface. Therefore, when a slot is reserved at an output port, the next slot must be reserved on the following output port along the path between a master and a slave module, and so on. When multiple connections are set up between several processing modules with timing guarantees, the slot allocation must be performed such that there are no clashes (i.e., there is no slot allocated to more than one connection). The slots must be reserved in such a way that data never has to contend with any other data. It is also called as contention free routing.
The task of finding an optimum slot allocation for a given network topology i.e. a given number of routers and network interfaces, and a set of connections between processing modules is a highly computational-intensive problem as it involves finding an optimal solution which requires exhaustive computation time.
An important feature for transmission of data between processing modules is the latency. A general definition of latency in networking could be summarized as the amount of time it takes a data packet to travel from source to destination. Together, latency and bandwidth define the speed and capacity of a network. The latency to access data depends on the size of such a slot table, assignment of slots for a given channel in the table and the burst size. The burst size is the amount of data that can be asked/sent in one request. When the number of slots allocated to a channel is less than the number of slots required to transfer a burst of data the latency to access data increases dramatically. In such case more than one revolution of the slot table is needed to completely send a burst of data. The waiting time for the slots that are not allocated to this connection is also added to the latency.
The network interfaces contain conventionally a queue per channel. The waiting time in that queue turns out to be the major contribution to the total communication latency. The larger the slot table in number of slots and the fewer slots are reserved for a channel, the higher the waiting latency.
The other problem is that when a single processing module requires many channels, say n, then the slot table requires at least n slots, one for each channel. However, this is not practical in general because the bandwidth requirements of the various channels may differ significantly which require even larger slot tables to allocate bandwidth at a finer granularity. The cost of the slot tables and thus of the network interfaces and thus of the network as a whole highly depends on the number of slots in the slot tables.
Therefore it is an object of the present invention is to provide an arrangement and a method having an improved slot allocation in a Network on Chip environment.
This object is solved by an integrated circuit according to claim 1 and to a method for time slot allocation according to claim 7.
It is proposed to share slots of channels having their origin at the same network interface. At least a part of the slots allocated to channels originating from the same network interface are shared. So a pool of slots is formed, which could be used by all channels together.
They will drastically reduce the latency. In particular the latency of channels having only a small number of slots allocated will be reduced. Since the number of slots in the slot table could be reduced by the sharing the memory space requirements in all network components are reduced.
Other aspects and advantages of the invention are defined in the dependent claims.
In a preferred embodiment of the invention all slots allocated to channels originating from the same network interface are shared. This will simplify the control of data transmission of the channels having shared slots.
In a further predetermined embodiment of the invention there is channel scheduler included in the network interface, the scheduler is provided for scheduling the data of the set of channels to the shared slots.
In a further predetermined embodiment of the invention the data of a channel are scheduled by the scheduler depending on the position in a queue. The control of the data transmission could be achieved by queuing the data belonging to set of channels in only one queue. Thus a first come first serve policy is implemented. This will further reduce the chip area required for the input queue in the network interface. Conventionally there is one queue per channel. According to the present invention it is advantageously to input all data of the shared channels in only one queue. The scheduler needs to schedule the data depending on its position in the queue.
In a preferred embodiment of the invention a scheduling of data of the set of channel is performed depending the filling status of the queue of the set of the channels. In an embodiment having a queue for each channels the scheduler will monitor the filling status of the queues of the channels. The first queue not being empty will be scheduled to be transferred. Then the scheduler will monitor the queues from that scheduled queue, wherein only queues are scheduled being not empty.
The invention also relates to a method for allocating time slots for data transmission in an integrated circuit having a plurality of processing modules and a network arranged for coupling the processing modules, and a plurality of network interfaces each being coupled between one of the processing modules and the network comprising the steps of: communicating between processing modules based on time division multiple access using time slots and contention free transmission by using channels; storing a slot table in each network interface including an allocation of a time slot to a certain channel, sharing of time slots allocated to channels originating from the same network interface.
The invention further relates to a data processing system comprising: a plurality of processing modules and a network arranged for coupling the processing modules, comprising: a network interface associated to each processing module which is provided for transmitting data to the network supplied by the associated processing module and for receiving data from the network destined for the associated processing module; wherein the data transmission between processing modules operates based on time division multiple access using time slots and contention free transmission by using a channels; each network interface includes a slot table for storing an allocation of a time slot to a certain channel, a sharing is provided of time slots allocated to channels originating from the same network interface.
Accordingly, the time slot allocation may also be performed in a multi-chip network or a system or network with several separate integrated circuits.
Preferred embodiments of the invention are described in detail below, by way of example only, with reference to the following schematic drawings.
The drawings are provided for illustrative purpose only and do not necessarily represent practical examples of the present invention to scale.
In the following the various exemplary embodiments of the invention are described.
Although the present invention is applicable in a broad variety of applications it will be described with the focus put on NoC, especially to AEthereal design. A further field for applying the invention might each NoC providing guaranteed services by using time slots and slot tables.
In the following the general architecture of a NoC will be described referring to
The embodiments relate to systems on chip SoC, i.e. a plurality of processing modules IP on the same chip communicate with each other via some kind of interconnect. The interconnect is embodied as a network on chip NoC. The network on chip NoC may include wires, bus, time-division multiplexing, switch, and/or routers within a network.
The underlying problem of high latency will be illustrated referring to
In the following the present invention will be explained referring to
The ten slots 0-9 allocated to the set are now designated by S. The ten slots S can be redistributed in the slot table ST. A good redistribution will place these slots S at equal distances in the slot table ST with possibly a minor over allocation of slots. This means that the ten slots S are located at slots 0, 4, 8, . . . , 36. However this distribution not only minimizes the worst case waiting time for a slot, but also allows to reduce the size of slot table by a factor of ten. This will cause a strong reduction of memory space required for the slot tables in each of the participating network components NI, R11-R44, etc. The reduced slot table ST has four slots only, and one of these slots 0-3 is assigned to channel set. A complete traversal of the small slot table is thus four slots, and the slot for channel set is thus available every four slots, which is the same as the example in which the ten slots were nicely distributed over the forty slots. Since all channels outgoing from the network interface NI are combined in that channel set, the rest of the slots in the slot table are used for channels not outgoing from the respective network interface NI.
When multiple channels a-d are combined into a channel set, some mechanism is required to schedule the data sequentially onto the network. There are basically two approaches for that. However before explaining the mechanism for scheduling the data of the multiple channels the structure of a network interface NI will be explained referring to
The NI receives the data at its input port 42 from the transmitting processing module IP. The NI outputs the packaged data at its output 43 to the router in form of a data sequence. The data to be transmitted are supplied to the queue 44. The first data in the queue 44 is monitored by the request generator 45. The request generator 45 detects the data and generates a request req_i based on the queue filling and the available remote space as stored in the remote space register 46. The request req_i for the queue is provided to the slot scheduler 55 for selecting the queue. The selection is may be performed by the slot scheduler 55 based on information from the slot table 54 and based on information of the used arbitration mechanism for controlling the set of channels. The scheduler 55 detects whether the data in the queue belongs to a channel a-d having shared slots or belonging to data which are not part of shared channel set slots. As soon as the queue is selected in the scheduler 55 it is provided to a unit 51 which increments the packet lengths and to the header insertion unit 52, which controls whether a header H needs to be inserted or not. Routing information like the addresses is stored in a configurable routing information register 47. The credit counter 49 is incremented when data is consumed in the output queue and is decremented when new headers H are sent with credit value incorporated in the headers H. The routing information from the routing information register 47 as well as the value of the credit counter 49 is forwarded to the header unit 48 and form part of the header H. The header unit 48 receives the credit value and routing info and outputs the header data to the output multiplexer 50. The output multiplexer 50 multiplexes the data provided by the selected queue and the header info hdr provided from the header unit 48. When a data package is sent out the packet length is reset.
As shown in
FCFS (first come first serve) policy and reduce the queuing cost significantly. The information that was used to control the de-multiplexer in the conventional architecture must now be queued in parallel to the data queue or in the same queue and increasing the word width of the queue. This control information reflects the channel ID in the channel set and is used to, e.g., select the path of the channel.
A further not illustrated mechanism could be that the scheduler 55 may use a first-come first-serve (FCFS) policy. When this policy is used the order in which the IP writes its data to the NI is queued. The first element in the queue 44 then indicates from which data queue the data may come. Note that the FCFS policy is a bit harder to use when the channel set is made from data coming from multiple IP blocks.
An alternative could be a simple round-robin (RR) scheduler that selects the first queue (the first from the previously selected queue) in the channel set that is non-empty.
One advantage of the method is that latency can be reduced significantly. In the example given, the worst case waiting time for a slot is reduced by a factor of ten. And the higher the ratio of the total bandwidth and the lowest bandwidth of a group of channels initiating from the same NI, the higher the latency reduction gets.
Another advantage is that this scheme does not require that all the channels in the set have both the same source and same destination. All that is required is that the channels have the same source.
Yet another advantage is that this scheme allows to reduce the size of the slot table. The example in this document shows a reduction of a factor of ten.
Yet another advantage is that this scheme allows to reduce the number of queues in the network interface. Referring to this example, one queue needs to be used instead of four.
Previous two advantages reduce the cost of the NI significantly, as the cost of the slot table and queues are dominant in the NI. Moreover, in practical networks it was further found that the cost of the NI dominates.
The only one disadvantage is that the more the channel set diverges, the more the overallocation of slots for the channels is required.
In systems in which the communication of data streams is done via shared memory the application of the invention is very important. In these schemes there are many processing modules writing and reading from a shared memory or multiple memories in general. It is typical of processing modules (CPUs) to have non-blocking writes and blocking reads. And hence the performance of the system depends highly on the latency of the reads. As the reads represents many data streams, all origination from the memory or memory controller the presented invention is very beneficial. Since there are many channels originating from the memory latency is reduced significantly, the slot-table size can be reduced significantly and the queue cost can be reduced significantly.
As all data streams go back and forth to memory, the overallocation is higher as one goes closer to the processing modules. But as all the streaming goes via memory, this overallocation is not a problem at all.
The invention is explained in the context of multiple synchronized TDMA however it is also applicable for single TDMA systems. In general it is applicable to interconnect structures basing on connections and providing guarantees.
It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design many alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word “comprising” does not exclude the presence of elements or steps other than those listed in a claim. The word “a” or “an” preceding an element does not exclude the presence of a plurality of such elements. In the device claim enumerating several means, several of these means can be embodied by one and the same item of hardware. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Furthermore, any reference signs in the claims shall not be construed as limiting the scope of the claims.
| Number | Date | Country | Kind |
|---|---|---|---|
| 05102702.7 | Apr 2005 | EP | regional |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/IB06/51012 | 4/4/2006 | WO | 00 | 10/5/2007 |
