}
	// Special case: -1 for either completions or parallelism means leave nil (negative is not allowed
	// in practice by validation.
	if completions >= 0 {
		j.Spec.Completions = &completions
	} else {
		j.Spec.Completions = nil
	}
	if parallelism >= 0 {
		j.Spec.Parallelism = &parallelism
	} else {
		j.Spec.Parallelism = nil
	}
	j.Spec.BackoffLimit = &backoffLimit

	return j
}

					Template:    validPodTemplateSpec,
					Parallelism: ptr.To[int32](4),
				},
				Status: batch.JobStatus{
					Active: 12,
				},
			},
		},
		"parallelism changes not allowed": {
			job: &batch.Job{
				ObjectMeta: validObjectMeta,
				Spec: batch.JobSpec{
					Selector:    validSelector,
					Template:    validPodTemplateSpec,
					Parallelism: ptr.To[int32](3),
				},
			},

					Completions:    pointer.Int32(2),
					Parallelism:    pointer.Int32(100000),
				},
			},
		},
		"valid parallelism and maxFailedIndexes for high completions when backoffLimitPerIndex is used": {
			job: batch.Job{
				ObjectMeta: validJobObjectMeta,
				Spec: batch.JobSpec{
					Completions:          pointer.Int32(100_000),
					Parallelism:          pointer.Int32(100_000),

  let summary = "Reorder replicated and partitioned input ops.";

  let description = [{
    This pass rewrites how data parallelism and model parallelism is expressed for
    inputs. It reorders `tf.TPUPartitionedInput` (model parallelism) and
    `tf.TPUReplicatedInput` (data parallelism) ops. It transforms a DAG where
    multiple `tf.TPUPartitionedInput` ops are feeding into a single

		i--
		dAtA[i] = 0x18
	}
	if m.Completions != nil {
		i = encodeVarintGenerated(dAtA, i, uint64(*m.Completions))
		i--
		dAtA[i] = 0x10
	}
	if m.Parallelism != nil {
		i = encodeVarintGenerated(dAtA, i, uint64(*m.Parallelism))
		i--
		dAtA[i] = 0x8
	}
	return len(dAtA) - i, nil
}

func (m *JobStatus) Marshal() (dAtA []byte, err error) {
	size := m.Size()

    %0 = "tf.B"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
    func.return %0 : tensor<?xi32>
  }
}

// -----

// Tests simple case of `tf_device.cluster_func` on TPU with replication. Under
// data parallelism replicated host devices are also added to the
// tf_device.replicate

		completions = fmt.Sprintf("%d/%d", obj.Status.Succeeded, *obj.Spec.Completions)
	} else {
		parallelism := int32(0)
		if obj.Spec.Parallelism != nil {
			parallelism = *obj.Spec.Parallelism
		}
		if parallelism > 1 {
			completions = fmt.Sprintf("%d/1 of %d", obj.Status.Succeeded, parallelism)
		} else {
			completions = fmt.Sprintf("%d/1", obj.Status.Succeeded)
		}
	}
	var jobDuration string

	const parallelism = 16

	db := newTestDB(b, "magicquery")
	defer closeDB(b, db)
	db.SetMaxIdleConns(runtime.GOMAXPROCS(0) * parallelism)

	stmt, err := db.Prepare("SELECT|magicquery|op|op=?,millis=?")
	if err != nil {
		b.Fatal(err)
	}
	defer stmt.Close()

	b.SetParallelism(parallelism)
	b.RunParallel(func(pb *testing.PB) {
		for pb.Next() {

func BenchmarkClientServerParallel(b *testing.B) {
	for _, parallelism := range []int{4, 64} {
		b.Run(fmt.Sprint(parallelism), func(b *testing.B) {
			run(b, func(b *testing.B, mode testMode) {
				benchmarkClientServerParallel(b, parallelism, mode)
			}, []testMode{http1Mode, https1Mode, http2Mode})
		})
	}
}

func benchmarkClientServerParallel(b *testing.B, parallelism int, mode testMode) {
	b.ReportAllocs()

				b.readySema <- true
			}
		}

		if a == root {
			close(b.readySema)
		}
	}

	var wg sync.WaitGroup

	// Kick off goroutines according to parallelism.
	// If we are using the -n flag (just printing commands)
	// drop the parallelism to 1, both to make the output
	// deterministic and because there is no real work anyway.
	par := cfg.BuildP
	if cfg.BuildN {
		par = 1
	}
	for i := 0; i < par; i++ {