CN110048966B

CN110048966B - Coflow scheduling method for minimizing system overhead based on deadline

Info

Publication number: CN110048966B
Application number: CN201910177871.3A
Authority: CN
Inventors: 李克秋; 王春晓; 周晓波; 徐仁海
Original assignee: Tianjin University
Current assignee: Tianjin University
Priority date: 2019-03-10
Filing date: 2019-03-10
Publication date: 2021-12-17
Anticipated expiration: 2039-03-10
Also published as: CN110048966A

Abstract

The invention relates to the technical field of computer networks and the field of data center networks, and aims to provide a flow scheduling method for minimizing system overhead based on deadline, which not only can complete more flows before the deadline is reached, but also designs a special scheduling method for the flows which miss the deadline, so as to minimize the system overhead caused by the flows missing the deadline. Therefore, the technical scheme adopted by the invention is that the Coflow scheduling method for minimizing the system overhead based on the Deadline comprises a Coflow scheduling step meeting the Deadline and a Coflow scheduling step missing the Deadline, wherein the Coflow is a set of parallel flows with correlation in a communication stage of parallel computation. The invention is mainly applied to network communication occasions.

Description

Coflow scheduling method for minimizing system overhead based on deadline

Technical Field

The invention relates to the technical field of computer networks and the field of data center networks. Particularly, under the data center network environment, research and development of a flow traffic scheduling technology based on deadline are carried out.

Background

The data center is an infrastructure for calculation, storage and data transmission, and integrates various software and hardware resources and key business systems. With the advent of cloud computing, big data, internet of things and artificial intelligence era, data traffic in a data center is increased explosively, the traditional data processing technology is not enough to meet the processing requirement of mass data, a parallel computing technology arises, and the mainstream distributed computing platforms at present are MapReduce, Spark, Dryad, TensorfLow and the like. The traffic transmission time of parallel computation in the big data processing process is larger and larger, for example, 33% of the running time of analysis work on network transmission is spent on Facebook. Network transmission gradually becomes a bottleneck of application performance, and therefore, how to optimize network transmission is a key problem of a data center network.

A conventional data center network transmission traffic scheduling technique is a single traffic (flow) based scheduling technique. The conventional scheduling technology attempts to distribute the demands of applications hosted on a data center by users to flow, schedule flows, decide the sequence of resources such as bandwidth allocated to the flows and the like, and perform resource allocation, so that it is hopeful that the demands of a network application layer can be transferred to a network transmission layer. However, a single flow is generally not able to take on the needs of the entire application. Often, an application has multiple flows in parallel with different demands (e.g., multiple flows have different deadlines demands), and thus a single flow cannot represent the demands of the entire application. The flow-based scheduling result may be that a certain flow of an application is scheduled and completed very early, but other flows of the application are completed later, resulting in the application eventually completing later, failing to meet the user's requirements and leaving a very poor user experience, or even the user no longer willing to host the application with the data center. In order to overcome the defect that flow scheduling cannot convey application requirements from an application layer to a transport layer, researchers propose a flow scheduling technology with a flow set as granularity, and a magpie bridge is built between the application layer requirement representation and a network transport layer. Flow is defined as the set of parallel flows (flows) with dependencies in the communication phase of parallel computation. For example, the Shuffle process between Map and Reduce of the Mapreduce framework is a flow, the data flow in the Shuffle process is parallel, and Reduce calculation can be started only after the last flow is finished transmission, that is, the Map phase is finished. The main characteristic of Coflow is that although multiple flows are parallel, as a whole, only the last flow transmission is completed, and the whole (Coflow) is really completed, i.e. the flow completion time (CCT) of Coflow depends on the latest completed flow. Since reducing the completion time of a single flow does not necessarily reduce the completion time of an application, and reducing the completion time of a Coflow can almost reduce the completion time of an application, how to schedule the Coflow becomes a core problem of current data center network transmission traffic scheduling.

The existing flow scheduling methods can be divided into two types, namely an information-sensible scheduling method and an information-imperceptible scheduling method. The optimization goal for Coflow is mainly two, one is to reduce the Coflow completion time, and the other is to have more Coflows completed before the deadline. Existing scheduling mechanisms are broadly divided into two in terms of making more Coflows meet the deadline: firstly, an admission control principle is used, and judgment is carried out before scheduling, for the Coflow which can meet the deadline, the scheduling is admitted and the Coflow is ensured not to be preempted by other Coflow, and for the Coflow which cannot meet the deadline, all data flows are abandoned or retransmitted; the second method is to divide the bandwidth into two parts by using the multiplexing principle, one part uses the priority sequence mode to schedule Coflows which can meet the deadline, and the other part uses the weighted average bandwidth allocation mode to schedule Coflows which can not meet the deadline. The above scheduling scheme does not take into account the sensitivity of different Coflows to deadlines, and there is no effective scheduling for Coflows that miss deadlines. At present, the existing technical research shows that every 100ms of network delay is increased, amazon sales income is reduced by 1%, and every 400ms of delay is increased, google search amount is reduced by 0.6%, so that under the background of the prior art, extra system overhead is brought by traffic exceeding the cut-off time. How to schedule the Coflows that miss the deadline is a crucial challenge in the Coflow scheduling problem.

In view of the foregoing background and motivation, a method for deadline-based flow traffic scheduling is presented herein: on one hand, under the condition of limited data center network resources (bandwidth), more Coflows are completed before the deadline is reached as much as possible; on the other hand, if the Coflows still cannot be completed before the Deadline, the system overhead of the type cofow exceeding the Deadline is considered, and the system overhead is minimized.

Disclosure of Invention

In order to overcome the defects of the prior art and solve the problem of flow scheduling with a deadline requirement, the invention aims to provide a deadline-based flow scheduling method for minimizing system overhead, so that more Coflows are completed before the deadline is reached as far as possible, a special scheduling method is designed for the Coflows missing the time limit, and the system overhead caused by the Coflow missing the time limit is minimized. Therefore, the technical scheme adopted by the invention is that the Coflow scheduling method for minimizing the system overhead based on the Deadline comprises a Coflow scheduling step meeting the Deadline and a Coflow scheduling step missing the Deadline, wherein the Coflow is a set of parallel flows with correlation in a communication stage of parallel computation;

(1) coflow scheduling procedure satisfying deadline

Acquiring related information in a network, firstly, judging whether the flow meets the cut-off time by adopting a dynamic method, if so, determining a scheduling sequence by using an Earliest-Deadline-First (EDF) method, namely, the flow which is closer to the Deadline is scheduled First, and after determining the scheduling sequence, allocating the bandwidth which is as small as possible to each flow to ensure that the flow is completed before the cut-off time;

(2) missed deadline flow scheduling step

Scheduling for the missed deadline Coflow is to schedule the Coflows missed by the deadline by using the residual bandwidth if the residual bandwidth in the network is still unused in the process of satisfying the deadline Coflow scheduling, and minimize the system overhead caused by the Coflows exceeding the deadline.

Step (1) is to make all flowss finish at that moment of deadline, and the specific steps are as follows:

1) acquiring network bandwidth information, wherein the network bandwidth information specifically comprises the residual bandwidth size of each uplink port and each downlink port; acquiring information of each flow, specifically including the number, the deadline, the number and the serial number of a source port, the number and the serial number of a destination port, and the data volume of each flow contained in the flow;

2) sequencing all the flow according to an EDF sequence, namely sequencing and scheduling according to the sequence of cutoff time from small to large;

3) judging whether the flow can be finished before the deadline, wherein the specific judgment mode is as follows: dividing the data volume of each flow in the flow by the deadline to obtain a corresponding expected bandwidth, and then judging whether the residual available bandwidth of each flow at the corresponding outlet and inlet meets the expected bandwidth or not;

4) if the residual available bandwidth of each flow of the flow is greater than the expected bandwidth at the corresponding destination port and source port, the flow can be completed before the deadline time, the bandwidth as small as possible is allocated to the flow, all flows are completed simultaneously with the flow completed at the latest, and the allocated bandwidth is the data volume of the flow divided by the deadline of the flow;

5) if only one flow exists in the flow, and the remaining available bandwidth of the corresponding exit or entrance is smaller than the expected bandwidth, the flow cannot be completed before the deadline, and the flow is put into a missed deadline set to wait for a subsequent missed deadline flow scheduling method to perform scheduling;

6) and (4) updating the residual bandwidth size of the network link, subtracting the bandwidth allocated in the step (4) and scheduling the next flow.

In the step (2), a cost function is specifically used for quantifying the relationship between the system overhead and the size exceeding the cut-off time, the cost function is set as a monotonically increasing linear function about that t is CCT-deadline, and the following specific steps are taken:

1) acquiring the respective cost function coefficients, namely slopes, of all the Coflows in the missed deadline set;

2) acquiring respective lengths of all Coflows in the missing deadline set, wherein the length of the cofow is the data volume of the maximum flow contained in the cofow;

3) calculating the skew of the Coflow in the missed cut-off time set, namely the ratio of the length of the Coflow to the cost function coefficient, and sequencing from small to large according to the skew;

4) and (4) sequentially allocating the network residual bandwidths according to the sequence of the step (3).

The invention has the characteristics and beneficial effects that:

the efficient scheduling of the data center flow is realized, more flows can be completed before the deadline, a special method is designed for scheduling the flow which misses the deadline, and the system overhead caused by the fact that the flow misses the deadline is minimized.

Description of the drawings:

FIG. 1 shows a flow network scheduling abstraction model and input cases.

FIG. 2 is an exemplary graph of the Cost function.

Fig. 3 shows the scheduling results (including total CCT and system overhead) for all possible scheduling orders.

Fig. 4 is a flowchart of a flow scheduling method for minimizing system overhead based on deadline according to the present invention.

Detailed Description

The invention aims to solve the problem of traffic scheduling with a deadline requirement. There are two main goals: one is that under the condition of limited data center network resources (bandwidth), more Coflows are completed before the deadline is reached as much as possible; secondly, a special scheduling scheme is designed for the Coflows exceeding the deadline, and the system overhead caused by the fact that the Coflows exceeds the deadline is minimized.

In order to overcome the defects of the prior art, the invention provides a Coflow scheduling method for minimizing system overhead based on deadline, which not only enables more Coflows to be completed before the deadline arrives as much as possible, but also designs a special scheduling method for the Coflows missing the time limit, and minimizes the system overhead caused by the Coflow missing the time limit. The scheduling method of the invention consists of a Coflow scheduling method meeting the deadline and a Coflow scheduling method missing the deadline.

1 Coflow scheduling method meeting deadline

Acquiring related information in a network, firstly, judging whether the flow meets the Deadline by adopting a dynamic method, and if so, determining a scheduling sequence by using an early-Deadline-First (EDF) method, namely, the flow closer to the Deadline is scheduled more First. After determining the scheduling order, we allocate as little bandwidth as possible to each flow to ensure that it is completed before the deadline. The completion time of the Coflow cannot be shortened quickly by a single flow transfer, and the flow with the latest completion time is determined, so that all flows of the Coflow can be completed at the same time with the flow with the latest completion time. For the method, we focus on making as many Coflows as possible complete at the deadline, so we can make all flows complete at that moment in the deadline, so that the bandwidth allocated to the cofow is minimal, the remaining available bandwidth in the network is as large as possible, and more Coflows can be scheduled to make them complete before the respective deadline. For ease of description, all Coflows in this method assume that time 0 arrives. The method comprises the following specific steps:

1) acquiring network bandwidth information, wherein the network bandwidth information specifically comprises the residual bandwidth size of each uplink port and each downlink port; the information of each flow that arrives is obtained, which specifically includes the number, deadline, the number and number of ingress ports (source ports), the number and number of egress ports (destination ports), and the data volume of each flow included in the flow.

2) And sequencing all the flow in an EDF sequence, namely sequencing and scheduling in a sequence with cutoff time from small to large.

3) Judging whether the flow can be finished before the deadline, wherein the specific judgment mode is as follows: dividing the data size of each flow in the flow by deadline (deadline) to obtain a corresponding expected bandwidth, and then judging whether the remaining available bandwidth of each flow at the corresponding outlet and inlet meets the expected bandwidth.

4) If each flow of the Coflow is greater than the expected bandwidth in the corresponding egress and ingress remaining available bandwidths, the Coflow can be completed before the deadline time, and the smallest possible bandwidth is allocated to the Coflow, so that all flows are completed simultaneously with the latest completed flow, and the allocated bandwidth is the data size of the flow divided by the deadline of the Coflow.

5) If only one flow exists in the flow, and the remaining available bandwidth of the corresponding exit or entry is smaller than the expected bandwidth, the flow cannot be completed before the deadline, and the flow is put into the missed deadline set to wait for a subsequent missed deadline flow scheduling method to perform scheduling.

2 missing deadline flow scheduling module

The deadline missing Coflow scheduling module is intended to schedule Coflows that miss deadlines, and in satisfying the deadline Coflow scheduling, the remaining bandwidth in the network is still unused, and the module uses the remaining bandwidth to schedule Coflows that miss deadlines and minimize the system overhead caused by Coflows that exceed the deadlines. As the flow exceeds the deadline time, the brought system overhead also becomes larger, in the method, the relationship between the system overhead and the size of the deadline time is quantified by a cost function, and for the sake of simplicity, the cost function is set as a monotonically increasing linear function about the condition that t is CCT-deadline. The method comprises the following specific steps:

1) the respective cost function coefficients (i.e., slopes) of all Coflows in the miss deadline set are obtained.

2) And acquiring respective lengths of all Coflows in the missing deadline set, wherein the length of the Coflow in the method is the data volume of the maximum flow contained in the Coflow.

3) And calculating the skew of the Coflow in the missing cut-off time set, namely the ratio of the length of the Coflow to the cost function coefficient, and sequencing from small to large according to the skew.

Examples of the present invention are described in further detail below with reference to the accompanying drawings.

FIG. 1 is an example of a data center network abstraction model and a multi-Coflow schedule for the Coflow scheduling problem. Without loss of generality, and similar to other important research works in China and abroad, we abstract the data center network into an abstract network that interconnects all servers by a non-blocking large switch as shown in fig. 1. The network throughput inside the abstract network is 100% and the bandwidth capacity is infinite, so the situation that multiple Coflows cause network congestion due to competition for bandwidth resources does not occur, but the bandwidth resources on each server port are limited, including an Ingress port (Ingress port) bound to the uplink and an Egress port (Egress port) bound to the downlink, and multiple Coflows arrive at the same time and will compete for bandwidth resources on these ports intensely, so that in the Coflow scheduling problem we only close the bandwidth allocation on the Ingress port and the Egress port. In the multiple Coflows example in this model, each ingress port has flows from one or more Coflows to the respective egress port. For ease of illustration, we put them in a virtual queue at the ingress port, and usually assume that the bandwidth capacity of each port is 1 unit, as shown, there are three Coflows in the network, and their detailed information is: coflow1 contains two flows, with data volumes of 6, 2; coflow2 contains three flows, with data volumes of 2,3, and 3; coflow3 contains three flows, all with a data volume of 2. Flow entry port of Coflow1 for data volume 6 is 2 and exit port is 1, and the detailed entry and exit port information for each of the remaining flows is shown in FIG. 1. Assuming that the deadlines of all three Coflows are 1 unit time, through calculation, none of the three Coflows can meet the requirement that the deadline is 1, and corresponding system overhead (cost) exceeding the deadline is brought, and we need to perform reasonable scheduling and minimize the system overhead.

Fig. 2 is a schematic diagram of the change of the Cost of each Colfow with the exceeding of the cutoff time, and it is assumed that the Cost functions corresponding to all Coflows are linear functions increasing monotonically, and it can be known from fig. 2 that the system overhead caused by the cofow exceeding the cutoff time increases as the exceeding of the cutoff time increases, and the Cost function coefficients of different Coflows are different, so that the system overhead caused by two Coflows with equal exceeding of the cutoff time is also different. The length and cost coefficients for the flow are shown in the following table in conjunction with fig. 1 and 2:

Coflow	cost coefficient (slope)	Flow length
			Coflow1(C1)	K1＝1	L1＝6
Coflow2(C2)	K2＝2	L2＝3
			Coflow3(C3)	K3＝0.5	L3＝2

Fig. 3 is a diagram of scheduling results and system overhead under different scheduling orders, in which all six scheduling orders of three Coflows and corresponding system overhead are listed, for example, fig. 3 (a) is used to describe how cost is calculated: the scheduling order of fig. 3 (a) is C1, C2, C3, and their completion times are 6, 8, and 10, respectively, simply calculated. Therefore, the overflow deadline of Coflow1 is 5, the overflow deadline of Coflow2 is 7, the overflow deadline of Coflow3 is 9, and the time is substituted into the cost function to obtain the system overhead caused by the Coflow missing the deadline, thereby obtaining the total system overhead. The total overhead for fig. 3 (a) is cost 5 × 1+7 × 2+9 × 0.5 × 23.5. The overhead calculation results for the remaining scheduling order are shown in fig. 3. As can be seen from comparison with other subgraphs in fig. 3, fig. 3 (f) shows that although the total completion time is minimum (CCT is 17), for Coflow for which the deadline has been missed, excessive pursuit of a small completion time increases the overhead (cost is 17.5). FIG. 3 (d) is the scheduling result obtained by the scheduling scheme proposed by the missing deadline Coflow scheduling module, i.e. sorting C2, C3, C1 (skew) according to the ratio (skew) of the length of all Coflows and the cost function coefficient from small to large₂<skew₃<skew₁) The flow scheduling is performed according to this sequence, and compared with other scheduling methods in fig. 3, the total overhead of the system is minimum.

Fig. 4 is a flowchart of a decode-based minimal system overhead flow scheduling method of the present invention, which specifically includes the following steps:

1) acquiring network bandwidth information, wherein the network bandwidth information specifically comprises the residual bandwidth size of each uplink port and each downlink port; and acquiring information of each flow, specifically including the number, arrival time, deadline, the number and number of ingress ports, the number and number of egress ports, and the data volume of each flow contained in the flow.

2) And sequencing all the flow according to the EDF sequence and storing the flow into a queue 1, namely sequencing and scheduling according to the sequence of the cutoff time from small to large.

3) And sequentially taking the flow at the head of the queue 1, and judging whether the flow can be completed before the deadline. The specific judgment method is as follows: and dividing the data volume of each flow in the flow by the deadline (deadline) to obtain a corresponding expected bandwidth, and then judging whether the residual available bandwidth of each flow at the corresponding outlet and inlet is greater than the expected bandwidth.

4) If the residual available Bandwidth of the inlet port and the outlet port corresponding to each flow is larger than the expected Bandwidth, namely Bandwidth_remain>size/deadline, the Coflow can be completed before the deadline, and each flow of the Coflow is allocated as little bandwidth as possible, i.e. all flows of the Coflow are completed at the moment of deadline. In the method, the deadline of Coflow n is T for example_nAllocating the least bandwidth to schedule the Coflow n, making the residual bandwidth as large as possible, so as to schedule more coflows, wherein, for a certain ingress port of the Coflow n, i is an egress port, j is an egress port, and j contains data volume of d_ijFlow of (1), allocated bandwidth of size d_ij/T_nAnd after the bandwidth is allocated, the residual bandwidth information on the i and j ports is updated.

5) Bandwidth if only one flow corresponding inlet port or outlet port has the residual available Bandwidth smaller than the expected Bandwidth_remain<size/deadline, the Coflow cannot meet the deadline, and the Coflow is put into a missed deadline set S to wait for a subsequent missed deadline Coflow scheduling method to schedule.

6) The scheduled cofow is removed from the head of queue 1.

7) And (5) circulating the steps (3) to (6) until no flow exists in the queue 1. At this point all Coflows that meet the deadline have completed scheduling.

8) Obtaining the cost function coefficients of all flow in the set S, wherein the cost functions of all flow are monotonously increasing linear functions, for example, the cost coefficient of Coflow1 is K₁。

9) And acquiring the lengths of all Coflows in the set S, wherein the Length of the Coflow in the method is the size of the maximum flow data contained in the Coflow, and for example, the Length (Length) of the Coflow1 is L1.

10) Calculating the skew L Length/K of Coflow in the missing deadline set S, sorting the Coflow according to the sequence of the skew from small to large, storing the Coflow into a queue 2, and allocating the bandwidth to the Coflow according to the sorting in the following step, wherein the skew L for the Coflow1₁/K₁。

11) And allocating the bandwidth to the cofow at the head of the queue 2, and allocating all the residual bandwidth of the current network to the cofow.

12) The flow removal that completed scheduling in the previous step.

13) And (5) repeating the steps (11) and (12) until the Coflow in the missed deadline aggregate finishes scheduling.

Claims

1. A flow scheduling method based on Deadline for minimizing system overhead is characterized by comprising a flow scheduling step meeting Deadline and a flow scheduling step missing Deadline, wherein the flow is a set of parallel flows with correlation in a communication stage of parallel computation;

(1) coflow scheduling procedure satisfying deadline

(2) missed deadline flow scheduling step

Scheduling for the missed deadline Coflow is that in the process of meeting the deadline Coflow scheduling, if the residual bandwidth in the network is still unused, the missed deadline Coflow is scheduled by using the residual bandwidth, and the system overhead caused by the Coflow exceeding the deadline is minimized;

in the step (2), a cost function is specifically used to quantify the relationship between the system overhead and the size exceeding the cutoff time, the cost function is set as a monotonically increasing linear function with respect to t ═ CCT-deadline, CCT represents the completion time of the flow, and the following specific steps are taken:

2. The flow scheduling method of claim 1, wherein the step (1) is to let all flows complete at the time of deadline, and comprises the following steps: