From: Tom Clegg <tom@curii.com>
Date: Tue, 8 Nov 2022 16:56:36 +0000 (-0500)
Subject: 19702: Increase default ReserveExtraRAM, allow keepstore overhead.
X-Git-Tag: 2.5.0~38^2~1
X-Git-Url: https://git.arvados.org/arvados.git/commitdiff_plain/5ee9b22d0777d7ef42fa6acc05b666fd30e1ab2b

19702: Increase default ReserveExtraRAM, allow keepstore overhead.

Arvados-DCO-1.1-Signed-off-by: Tom Clegg <tom@curii.com>
---

diff --git a/lib/config/config.default.yml b/lib/config/config.default.yml
index 09c068a0b9..0246cb88d5 100644
--- a/lib/config/config.default.yml
+++ b/lib/config/config.default.yml
@@ -1028,7 +1028,7 @@ Clusters:
 
       # Extra RAM to reserve on the node, in addition to
       # the amount specified in the container's RuntimeConstraints
-      ReserveExtraRAM: 256MiB
+      ReserveExtraRAM: 550MiB
 
       # Minimum time between two attempts to run the same container
       MinRetryPeriod: 0s
diff --git a/lib/dispatchcloud/node_size.go b/lib/dispatchcloud/node_size.go
index 7c7643bfc7..8cc63dc208 100644
--- a/lib/dispatchcloud/node_size.go
+++ b/lib/dispatchcloud/node_size.go
@@ -110,7 +110,12 @@ func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvad
 
 	needRAM := ctr.RuntimeConstraints.RAM + ctr.RuntimeConstraints.KeepCacheRAM
 	needRAM += int64(cc.Containers.ReserveExtraRAM)
-	needRAM += int64(cc.Containers.LocalKeepBlobBuffersPerVCPU * needVCPUs * (1 << 26))
+	if cc.Containers.LocalKeepBlobBuffersPerVCPU > 0 {
+		// + 200 MiB for keepstore process + 10% for GOGC=10
+		needRAM += 220 << 20
+		// + 64 MiB for each blob buffer + 10% for GOGC=10
+		needRAM += int64(cc.Containers.LocalKeepBlobBuffersPerVCPU * needVCPUs * (1 << 26) * 11 / 10)
+	}
 	needRAM = (needRAM * 100) / int64(100-discountConfiguredRAMPercent)
 
 	ok := false
diff --git a/lib/dispatchcloud/node_size_test.go b/lib/dispatchcloud/node_size_test.go
index eb3648e8ac..86bfbec7b6 100644
--- a/lib/dispatchcloud/node_size_test.go
+++ b/lib/dispatchcloud/node_size_test.go
@@ -80,7 +80,10 @@ func (*NodeSizeSuite) TestChoose(c *check.C) {
 			"costly": {Price: 4.4, RAM: 4000000000, VCPUs: 8, Scratch: 2 * GiB, Name: "costly"},
 		},
 	} {
-		best, err := ChooseInstanceType(&arvados.Cluster{InstanceTypes: menu, Containers: arvados.ContainersConfig{ReserveExtraRAM: 268435456}}, &arvados.Container{
+		best, err := ChooseInstanceType(&arvados.Cluster{InstanceTypes: menu, Containers: arvados.ContainersConfig{
+			LocalKeepBlobBuffersPerVCPU: 1,
+			ReserveExtraRAM:             268435456,
+		}}, &arvados.Container{
 			Mounts: map[string]arvados.Mount{
 				"/tmp": {Kind: "tmp", Capacity: 2 * int64(GiB)},
 			},
@@ -98,7 +101,30 @@ func (*NodeSizeSuite) TestChoose(c *check.C) {
 	}
 }
 
-func (*NodeSizeSuite) TestChoosePreemptable(c *check.C) {
+func (*NodeSizeSuite) TestChooseWithBlobBuffersOverhead(c *check.C) {
+	menu := map[string]arvados.InstanceType{
+		"nearly": {Price: 2.2, RAM: 4000000000, VCPUs: 4, Scratch: 2 * GiB, Name: "small"},
+		"best":   {Price: 3.3, RAM: 8000000000, VCPUs: 4, Scratch: 2 * GiB, Name: "best"},
+		"costly": {Price: 4.4, RAM: 12000000000, VCPUs: 8, Scratch: 2 * GiB, Name: "costly"},
+	}
+	best, err := ChooseInstanceType(&arvados.Cluster{InstanceTypes: menu, Containers: arvados.ContainersConfig{
+		LocalKeepBlobBuffersPerVCPU: 16, // 1 GiB per vcpu => 2 GiB
+		ReserveExtraRAM:             268435456,
+	}}, &arvados.Container{
+		Mounts: map[string]arvados.Mount{
+			"/tmp": {Kind: "tmp", Capacity: 2 * int64(GiB)},
+		},
+		RuntimeConstraints: arvados.RuntimeConstraints{
+			VCPUs:        2,
+			RAM:          987654321,
+			KeepCacheRAM: 123456789,
+		},
+	})
+	c.Check(err, check.IsNil)
+	c.Check(best.Name, check.Equals, "best")
+}
+
+func (*NodeSizeSuite) TestChoosePreemptible(c *check.C) {
 	menu := map[string]arvados.InstanceType{
 		"costly":      {Price: 4.4, RAM: 4000000000, VCPUs: 8, Scratch: 2 * GiB, Preemptible: true, Name: "costly"},
 		"almost best": {Price: 2.2, RAM: 2000000000, VCPUs: 4, Scratch: 2 * GiB, Name: "almost best"},