Mesos源码分析(5): Mesos Master的启动之四

 

5. Create an instance of allocator.

 

代码如下

 

Mesos源码中默认的Allocator,即HierarchicalDRFAllocator的位置在$MESOS_HOME/src/master/allocator/mesos/hierarchical.hpp,而DRF中对每个Framework排序的Sorter位于$MESOS_HOME/src/master/allocator/sorter/drf/sorter.cpp,可以查看其源码了解它的工作原理。

 

HierarchicalDRF的基本原理

 

如何作出offer分配的决定是由资源分配模块Allocator实现的,该模块存在于Master之中。资源分配模块确定Framework接受offer的顺序,与此同时,确保在资源利用最大化的条件下公平地共享资源。

由于Mesos为跨数据中心调度资源并且是异构的资源需求时,资源分配相比普通调度将会更加困难。因此Mesos采用了DRF(主导资源公平算法 Dominant Resource Fairness)

Framework拥有的全部资源类型份额中占最高百分比的就是Framework的主导份额。DRF算法会使用所有已注册的Framework来计算主导份额,以确保每个Framework能接收到其主导资源的公平份额。

 

举个例子

 

考虑一个9CPU,18GBRAM的系统,拥有两个用户,其中用户A运行的任务的需求向量为{1CPU, 4GB},用户B运行的任务的需求向量为{3CPU,1GB},用户可以执行尽量多的任务来使用系统的资源。

在上述方案中,A的每个任务消耗总cpu的1/9和总内存的2/9,所以A的dominant resource是内存;B的每个任务消耗总cpu的1/3和总内存的1/18,所以B的dominant resource为CPU。DRF会均衡用户的dominant shares,执行3个用户A的任务,执行2个用户B的任务。三个用户A的任务总共消耗了{3CPU,12GB},两个用户B的任务总共消耗了{6CPU,2GB};在这个分配中,每一个用户的dominant share是相等的,用户A获得了2/3的RAM,而用户B获得了2/3的CPU。

以上的这个分配可以用如下方式计算出来:x和y分别是用户A和用户B的分配任务的数目,那么用户A消耗了{xCPU,4xGB},用户B消耗了{3yCPU,yGB},在图三中用户A和用户B消耗了同等dominant resource;用户A的dominant share为4x/18,用户B的dominant share为3y/9。所以DRF分配可以通过求解以下的优化问题来得到:

 

max(x,y)     #(Maximize allocations)

    subject to

        x + 3y <= 9         #(CPU constraint)

        4x + y <= 18         #(Memory Constraint)

            2x/9 = y/3     #(Equalize dominant shares)

 

最后解出x=3以及y=2,因而用户A获得{3CPU,12GB},B得到{6CPU, 2GB}。

 

HierarchicalDRF核心算法实现

 

HierachicalDRF的实现在Src/main/allocator/mesos/hierarchical.cpp中

 

 

不是每次把所有的资源都给所有的framework,而是根据资源分配算法,每个framework拿到的不同

 

  1. void HierarchicalAllocatorProcess::allocate(
  2.     const hashset<SlaveID>& slaveIds_)
  3. {
  4.   ++metrics.allocation_runs;
  5.  
  6.   // Compute the offerable resources, per framework:
  7.   // (1) For reserved resources on the slave, allocate these to a
  8.   // framework having the corresponding role.
  9.   // (2) For unreserved resources on the slave, allocate these
  10.   // to a framework of any role.
  11.   hashmap<FrameworkID, hashmap<SlaveID, Resources>> offerable;
  12.  
  13.   // NOTE: This function can operate on a small subset of slaves, we have to
  14.   // make sure that we don't assume cluster knowledge when summing resources
  15.   // from that set.
  16.  
  17.   vector<SlaveID> slaveIds;
  18.   slaveIds.reserve(slaveIds_.size());
  19.  
  20.   // Filter out non-whitelisted and deactivated slaves in order not to send
  21.   // offers for them.
  22.   foreach (const SlaveID& slaveId, slaveIds_) {
  23.     if (isWhitelisted(slaveId) && slaves[slaveId].activated) {
  24.       slaveIds.push_back(slaveId);
  25.     }
  26.   }
  27.  
  28.   // Randomize the order in which slaves' resources are allocated.
  29.   //
  30.   // TODO(vinod): Implement a smarter sorting algorithm.
  31.   std::random_shuffle(slaveIds.begin(), slaveIds.end());
  32.  
  33.   // Returns the __quantity__ of resources allocated to a quota role. Since we
  34.   // account for reservations and persistent volumes toward quota, we strip
  35.   // reservation and persistent volume related information for comparability.
  36.   // The result is used to determine whether a role's quota is satisfied, and
  37.   // also to determine how many resources the role would need in order to meet
  38.   // its quota.
  39.   //
  40.   // NOTE: Revocable resources are excluded in `quotaRoleSorter`.
  41.   auto getQuotaRoleAllocatedResources = [this](const string& role) {
  42.     CHECK(quotas.contains(role));
  43.  
  44.     // NOTE: `allocationScalarQuantities` omits dynamic reservation and
  45.     // persistent volume info, but we additionally strip `role` here.
  46.     Resources resources;
  47.  
  48.     foreach (Resource resource,
  49.              quotaRoleSorter->allocationScalarQuantities(role)) {
  50.       CHECK(!resource.has_reservation());
  51.       CHECK(!resource.has_disk());
  52.  
  53.       resource.set_role("*");
  54.       resources += resource;
  55.     }
  56.  
  57.     return resources;
  58.   };
  59.  
  60.   // Quota comes first and fair share second. Here we process only those
  61.   // roles, for which quota is set (quota'ed roles). Such roles form a
  62.   // special allocation group with a dedicated sorter.
  63.   foreach (const SlaveID& slaveId, slaveIds) {
  64.     foreach (const string& role, quotaRoleSorter->sort()) {
  65.       CHECK(quotas.contains(role));
  66.  
  67.       // If there are no active frameworks in this role, we do not
  68.       // need to do any allocations for this role.
  69.       if (!activeRoles.contains(role)) {
  70.         continue;
  71.       }
  72.  
  73.       // Get the total quantity of resources allocated to a quota role. The
  74.       // value omits role, reservation, and persistence info.
  75.       Resources roleConsumedResources = getQuotaRoleAllocatedResources(role);
  76.  
  77.       // If quota for the role is satisfied, we do not need to do any further
  78.       // allocations for this role, at least at this stage.
  79.       //
  80.       // TODO(alexr): Skipping satisfied roles is pessimistic. Better
  81.       // alternatives are:
  82.       // * A custom sorter that is aware of quotas and sorts accordingly.
  83.       // * Removing satisfied roles from the sorter.
  84.       if (roleConsumedResources.contains(quotas[role].info.guarantee())) {
  85.         continue;
  86.       }
  87.  
  88.       // Fetch frameworks according to their fair share.
  89.       foreach (const string& frameworkId_, frameworkSorters[role]->sort()) {
  90.         FrameworkID frameworkId;
  91.         frameworkId.set_value(frameworkId_);
  92.  
  93.         // If the framework has suppressed offers, ignore. The unallocated
  94.         // part of the quota will not be allocated to other roles.
  95.         if (frameworks[frameworkId].suppressed) {
  96.           continue;
  97.         }
  98.  
  99.         // Only offer resources from slaves that have GPUs to
  100.         // frameworks that are capable of receiving GPUs.
  101.         // See MESOS-5634.
  102.         if (!frameworks[frameworkId].gpuAware &&
  103.             slaves[slaveId].total.gpus().getOrElse(0) > 0) {
  104.           continue;
  105.         }
  106.  
  107.         // Calculate the currently available resources on the slave.
  108.         Resources available = slaves[slaveId].total - slaves[slaveId].allocated;
  109.  
  110.         // The resources we offer are the unreserved resources as well as the
  111.         // reserved resources for this particular role. This is necessary to
  112.         // ensure that we don't offer resources that are reserved for another
  113.         // role.
  114.         //
  115.         // NOTE: Currently, frameworks are allowed to have '*' role.
  116.         // Calling reserved('*') returns an empty Resources object.
  117.         //
  118.         // Quota is satisfied from the available non-revocable resources on the
  119.         // agent. It's important that we include reserved resources here since
  120.         // reserved resources are accounted towards the quota guarantee. If we
  121.         // were to rely on stage 2 to offer them out, they would not be checked
  122.         // against the quota guarantee.
  123.         Resources resources =
  124.           (available.unreserved() + available.reserved(role)).nonRevocable();
  125.  
  126.         // It is safe to break here, because all frameworks under a role would
  127.         // consider the same resources, so in case we don't have allocatable
  128.         // resources, we don't have to check for other frameworks under the
  129.         // same role. We only break out of the innermost loop, so the next step
  130.         // will use the same `slaveId`, but a different role.
  131.         //
  132.         // NOTE: The resources may not be allocatable here, but they can be
  133.         // accepted by one of the frameworks during the second allocation
  134.         // stage.
  135.         if (!allocatable(resources)) {
  136.           break;
  137.         }
  138.  
  139.         // If the framework filters these resources, ignore. The unallocated
  140.         // part of the quota will not be allocated to other roles.
  141.         if (isFiltered(frameworkId, slaveId, resources)) {
  142.           continue;
  143.         }
  144.  
  145.         VLOG(2) << "Allocating " << resources << " on agent " << slaveId
  146.                 << " to framework " << frameworkId
  147.                 << " as part of its role quota";
  148.  
  149.         // NOTE: We perform "coarse-grained" allocation for quota'ed
  150.         // resources, which may lead to overcommitment of resources beyond
  151.         // quota. This is fine since quota currently represents a guarantee.
  152.         offerable[frameworkId][slaveId] += resources;
  153.         slaves[slaveId].allocated += resources;
  154.  
  155.         // Resources allocated as part of the quota count towards the
  156.         // role's and the framework's fair share.
  157.         //
  158.         // NOTE: Revocable resources have already been excluded.
  159.         frameworkSorters[role]->add(slaveId, resources);
  160.         frameworkSorters[role]->allocated(frameworkId_, slaveId, resources);
  161.         roleSorter->allocated(role, slaveId, resources);
  162.         quotaRoleSorter->allocated(role, slaveId, resources);
  163.       }
  164.     }
  165.   }
  166.  
  167.   // Calculate the total quantity of scalar resources (including revocable
  168.   // and reserved) that are available for allocation in the next round. We
  169.   // need this in order to ensure we do not over-allocate resources during
  170.   // the second stage.
  171.   //
  172.   // For performance reasons (MESOS-4833), this omits information about
  173.   // dynamic reservations or persistent volumes in the resources.
  174.   //
  175.   // NOTE: We use total cluster resources, and not just those based on the
  176.   // agents participating in the current allocation (i.e. provided as an
  177.   // argument to the `allocate()` call) so that frameworks in roles without
  178.   // quota are not unnecessarily deprived of resources.
  179.   Resources remainingClusterResources = roleSorter->totalScalarQuantities();
  180.   foreachkey (const string& role, activeRoles) {
  181.     remainingClusterResources -= roleSorter->allocationScalarQuantities(role);
  182.   }
  183.  
  184.   // Frameworks in a quota'ed role may temporarily reject resources by
  185.   // filtering or suppressing offers. Hence quotas may not be fully allocated.
  186.   Resources unallocatedQuotaResources;
  187.   foreachpair (const string& name, const Quota& quota, quotas) {
  188.     // Compute the amount of quota that the role does not have allocated.
  189.     //
  190.     // NOTE: Revocable resources are excluded in `quotaRoleSorter`.
  191.     // NOTE: Only scalars are considered for quota.
  192.     Resources allocated = getQuotaRoleAllocatedResources(name);
  193.     const Resources required = quota.info.guarantee();
  194.     unallocatedQuotaResources += (required - allocated);
  195.   }
  196.  
  197.   // Determine how many resources we may allocate during the next stage.
  198.   //
  199.   // NOTE: Resources for quota allocations are already accounted in
  200.   // `remainingClusterResources`.
  201.   remainingClusterResources -= unallocatedQuotaResources;
  202.  
  203.   // To ensure we do not over-allocate resources during the second stage
  204.   // with all frameworks, we use 2 stopping criteria:
  205.   // * No available resources for the second stage left, i.e.
  206.   // `remainingClusterResources` - `allocatedStage2` is empty.
  207.   // * A potential offer will force the second stage to use more resources
  208.   // than available, i.e. `remainingClusterResources` does not contain
  209.   // (`allocatedStage2` + potential offer). In this case we skip this
  210.   // agent and continue to the next one.
  211.   //
  212.   // NOTE: Like `remainingClusterResources`, `allocatedStage2` omits
  213.   // information about dynamic reservations and persistent volumes for
  214.   // performance reasons. This invariant is preserved because we only add
  215.   // resources to it that have also had this metadata stripped from them
  216.   // (typically by using `Resources::createStrippedScalarQuantity`).
  217.   Resources allocatedStage2;
  218.  
  219.   // At this point resources for quotas are allocated or accounted for.
  220.   // Proceed with allocating the remaining free pool.
  221.   foreach (const SlaveID& slaveId, slaveIds) {
  222.     // If there are no resources available for the second stage, stop.
  223.     if (!allocatable(remainingClusterResources - allocatedStage2)) {
  224.       break;
  225.     }
  226.  
  227.     foreach (const string& role, roleSorter->sort()) {
  228.       foreach (const string& frameworkId_,
  229.                frameworkSorters[role]->sort()) {
  230.         FrameworkID frameworkId;
  231.         frameworkId.set_value(frameworkId_);
  232.  
  233.         // If the framework has suppressed offers, ignore.
  234.         if (frameworks[frameworkId].suppressed) {
  235.           continue;
  236.         }
  237.  
  238.         // Only offer resources from slaves that have GPUs to
  239.         // frameworks that are capable of receiving GPUs.
  240.         // See MESOS-5634.
  241.         if (!frameworks[frameworkId].gpuAware &&
  242.             slaves[slaveId].total.gpus().getOrElse(0) > 0) {
  243.           continue;
  244.         }
  245.  
  246.         // Calculate the currently available resources on the slave.
  247.         Resources available = slaves[slaveId].total - slaves[slaveId].allocated;
  248.  
  249.         // The resources we offer are the unreserved resources as well as the
  250.         // reserved resources for this particular role. This is necessary to
  251.         // ensure that we don't offer resources that are reserved for another
  252.         // role.
  253.         //
  254.         // NOTE: Currently, frameworks are allowed to have '*' role.
  255.         // Calling reserved('*') returns an empty Resources object.
  256.         //
  257.         // NOTE: We do not offer roles with quota any more non-revocable
  258.         // resources once their quota is satisfied. However, note that this is
  259.         // not strictly true due to the coarse-grained nature (per agent) of the
  260.         // allocation algorithm in stage 1.
  261.         //
  262.         // TODO(mpark): Offer unreserved resources as revocable beyond quota.
  263.         Resources resources = available.reserved(role);
  264.         if (!quotas.contains(role)) {
  265.           resources += available.unreserved();
  266.         }
  267.  
  268.         // It is safe to break here, because all frameworks under a role would
  269.         // consider the same resources, so in case we don't have allocatable
  270.         // resources, we don't have to check for other frameworks under the
  271.         // same role. We only break out of the innermost loop, so the next step
  272.         // will use the same slaveId, but a different role.
  273.         //
  274.         // The difference to the second `allocatable` check is that here we also
  275.         // check for revocable resources, which can be disabled on a per frame-
  276.         // work basis, which requires us to go through all frameworks in case we
  277.         // have allocatable revocable resources.
  278.         if (!allocatable(resources)) {
  279.           break;
  280.         }
  281.  
  282.         // Remove revocable resources if the framework has not opted for them.
  283.         if (!frameworks[frameworkId].revocable) {
  284.           resources = resources.nonRevocable();
  285.         }
  286.  
  287.         // If the resources are not allocatable, ignore. We can not break
  288.         // here, because another framework under the same role could accept
  289.         // revocable resources and breaking would skip all other frameworks.
  290.         if (!allocatable(resources)) {
  291.           continue;
  292.         }
  293.  
  294.         // If the framework filters these resources, ignore.
  295.         if (isFiltered(frameworkId, slaveId, resources)) {
  296.           continue;
  297.         }
  298.  
  299.         // If the offer generated by `resources` would force the second
  300.         // stage to use more than `remainingClusterResources`, move along.
  301.         // We do not terminate early, as offers generated further in the
  302.         // loop may be small enough to fit within `remainingClusterResources`.
  303.         const Resources scalarQuantity =
  304.           resources.createStrippedScalarQuantity();
  305.  
  306.         if (!remainingClusterResources.contains(
  307.                 allocatedStage2 + scalarQuantity)) {
  308.           continue;
  309.         }
  310.  
  311.         VLOG(2) << "Allocating " << resources << " on agent " << slaveId
  312.                 << " to framework " << frameworkId;
  313.  
  314.         // NOTE: We perform "coarse-grained" allocation, meaning that we always
  315.         // allocate the entire remaining slave resources to a single framework.
  316.         //
  317.         // NOTE: We may have already allocated some resources on the current
  318.         // agent as part of quota.
  319.         offerable[frameworkId][slaveId] += resources;
  320.         allocatedStage2 += scalarQuantity;
  321.         slaves[slaveId].allocated += resources;
  322.  
  323.         frameworkSorters[role]->add(slaveId, resources);
  324.         frameworkSorters[role]->allocated(frameworkId_, slaveId, resources);
  325.         roleSorter->allocated(role, slaveId, resources);
  326.  
  327.         if (quotas.contains(role)) {
  328.           // See comment at `quotaRoleSorter` declaration regarding
  329.           // non-revocable.
  330.           quotaRoleSorter->allocated(role, slaveId, resources.nonRevocable());
  331.         }
  332.       }
  333.     }
  334.   }
  335.  
  336.   if (offerable.empty()) {
  337.     VLOG(1) << "No allocations performed";
  338.   } else {
  339.     // Now offer the resources to each framework.
  340.     foreachkey (const FrameworkID& frameworkId, offerable) {
  341.       offerCallback(frameworkId, offerable[frameworkId]);
  342.     }
  343.   }
  344.  
  345.   // NOTE: For now, we implement maintenance inverse offers within the
  346.   // allocator. We leverage the existing timer/cycle of offers to also do any
  347.   // "deallocation" (inverse offers) necessary to satisfy maintenance needs.
  348.   deallocate(slaveIds_);
  349. }

 

上面这段算法非常复杂,概况来说调用了三个Sorter,对所有的Framework进行排序,哪个先得到资源,哪个后得到资源。

 

在src/master/allocator/mesos/hierarchical.hpp中,有对三个重要的sorter的定义和注释,可以帮助了解sorter的逻辑。

 

 

 

 

 

总的来说分两大步:

  • 先保证有quota的role
  • 然后其他的资源没有quota的再分

在每一步Hierachical的意思是两层排序

  • 一层是按照role排序
  • 第二层是相同的role的不同Framework排序

每一层的排序都是按照计算的share进行排序来先给谁,再给谁

 

在src/master/allocator/sorter/drf/sorter.cpp中

 

 

  1. double DRFSorter::calculateShare(const string& name)
  2. {
  3.   double share = 0.0;
  4.  
  5.   // TODO(benh): This implementation of "dominant resource fairness"
  6.   // currently does not take into account resources that are not
  7.   // scalars.
  8.  
  9.   foreach (const string& scalar, total_.scalarQuantities.names()) {
  10.     // Filter out the resources excluded from fair sharing.
  11.     if (fairnessExcludeResourceNames.isSome() &&
  12.         fairnessExcludeResourceNames->count(scalar) > 0) {
  13.       continue;
  14.     }
  15.  
  16.     // We collect the scalar accumulated total value from the
  17.     // `Resources` object.
  18.     //
  19.     // NOTE: Although in principle scalar resources may be spread
  20.     // across multiple `Resource` objects (e.g., persistent volumes),
  21.     // we currently strip persistence and reservation metadata from
  22.     // the resources in `scalarQuantities`.
  23.     Option<Value::Scalar> __total =
  24.       total_.scalarQuantities.get<Value::Scalar>(scalar);
  25.  
  26.     CHECK_SOME(__total);
  27.     const double _total = __total.get().value();
  28.  
  29.     if (_total > 0.0) {
  30.       double allocation = 0.0;
  31.  
  32.       // We collect the scalar accumulated allocation value from the
  33.       // `Resources` object.
  34.       //
  35.       // NOTE: Although in principle scalar resources may be spread
  36.       // across multiple `Resource` objects (e.g., persistent volumes),
  37.       // we currently strip persistence and reservation metadata from
  38.       // the resources in `scalarQuantities`.
  39.       Option<Value::Scalar> _allocation =
  40.         allocations[name].scalarQuantities.get<Value::Scalar>(scalar);
  41.  
  42.       if (_allocation.isSome()) {
  43.         allocation = _allocation.get().value();
  44.       }
  45.  
  46.       share = std::max(share, allocation / _total);
  47.     }
  48.   }
  49.  
  50.   return share / weights[name];
  51. }

 

 

Quota, Reservation, Role, Weight

 

  • 每个Framework可以有Role,既用于权限,也用于资源分配
  • 可以给某个role在offerResources的时候回复Offer::Operation::RESERVE,来预订某台slave上面的资源。Reservation是很具体的,具体到哪台机器的多少资源属于哪个Role
  • Quota是每个Role的最小保证量,但是不具体到某个节点,而是在整个集群中保证有这么多就行了。
  • Reserved资源也算在Quota里面。
  • 不同的Role之间可以有Weight

 

最后将resource交给每一个Framework

在allocate函数的最后,依次调用offerCallback来讲resource分配给每一个Framework

 

 

那offerCallback函数是什么时候注册进来的呢?

 

 

 

在Allocator的initialize函数中,OfferCallback被注册尽量,并且没过一段时间执行一次。

 

在Allocator初始化的时候,最后定义每allocationInterval运行一次

offerCallback是注册进来的函数,请记住。

 

6. flags.registry == "in_memory" or flags.registry == "replicated_log" 信息存储在内存,zk,本地文件夹

 

7. 选举和检测谁是Leader的对象初始化

 

Try<MasterContender*> contender_ = MasterContender::create(zk, flags.master_contender);

Try<MasterDetector*> detector_ = MasterDetector::create(zk, flags.master_detector);

 

8. 生成Master对象,启动Master线程

 

posted @ 2016-07-24 20:24  popsuper1982  阅读(1119)  评论(0编辑  收藏  举报