parsec is a DHT lookup performance measurement tool. It specifically measures the PUT and GET performance of the
IPFS public DHT but could also be configured to measure
other libp2p-kad-dht networks.
The setup is split into two components: a scheduler and a server.
The server is just a normal libp2p peer that supports and participates in the public IPFS DHT and exposes a lean HTTP API that allows the scheduler to issue publication and retrieval operations. Currently, in ProbeLab's deployment, the scheduler goes around all seven server nodes, instructs one to publish provider records for a random data blob and asks the other six to look them up. All seven servers take timing measurements about the publication or retrieval latencies and report back the results to the scheduler. The scheduler then tracks this information in a database for later analysis.
Next to the concept of servers and schedulers there's the concept of a fleet. A fleet is a set of server nodes that
have a common configuration. For example, we are running three different fleets with seven nodes each (in different regions): 1) default 2) optprov 3) fullrt.
Each of these three fleets are configured differently. The default fleet uses the default configuration in the go-libp2p-kad-dht repository, the optprov fleet uses the optimistic provide configuration to publish data into the DHT, and the fullrt fleet uses the accelerated DHT client.
Schedulers are then configured to interface with any combination of fleets. Right now, we have one scheduler for each fleet. As said above, it asks one node to publish content, then instructs the others to find the provider records, and then repeats the process with the next peer. However,
we could configure a scheduler that does the same thing but with nodes from multiple fleets e.g., default fullrt to check if content that's published with one implementation is reachable with another one.
You can run
docker compose upto start two servers and one scheduler and see them interact.
Right now, the server component is implemented in Go and uses the go-libp2p-kad-dht implementation. It consequently measures the Go implementations performance. However, other implementations exist that support the DHT protocol. These implementations can be easily integrated with this measurement infrastructure. They just need to behave as a parsec server. The existing schedulers can then be reused.
Here are the things that a new implementation would need to do:
- Expose an HTTP interface with three endpoints
- Upon startup write general information about the node configuration to a postgres database.
- Regularly refresh the heartbeat field in the database.
You can find the OpenAPI specification in the ./server.yaml.
The new server would need to initialize a postgres client. The default environment variables to configure the client are as follows:
PARSEC_DATABASE_HOST
PARSEC_DATABASE_PORT
PARSEC_DATABASE_NAME
PARSEC_DATABASE_PASSWORD
PARSEC_DATABASE_USER
PARSEC_DATABASE_SSL_MODEThen upon startup, the server needs to write a row into the nodes_ecs table. The definition looks like this:
CREATE TABLE nodes_ecs
(
-- auto generated, doesn't need to be set manually
id INT GENERATED ALWAYS AS IDENTITY,
-- available CPUs
cpu INT NOT NULL,
-- available memory rounded to the nearest MB
memory INT NOT NULL,
-- the peer ID of the libp2p host
peer_id TEXT NOT NULL,
-- in which region does this server/node run? Given via the AWS_REGION environment var
region TEXT NOT NULL,
-- os.Args - with which arguments was this server run?
cmd TEXT NOT NULL,
-- a fleet identifier (see section `Concepts` above)
fleet TEXT NOT NULL,
-- a JSON document with no enforced schema. Could be anything really. It's intended
-- to give information about the exact dependencies, and especially kad-dht implementation
-- that the server uses.
dependencies JSONB NOT NULL,
-- the private IP address of the server. The scheduler will query this table and use this
-- ip address to contact the HTTP API. In the ECS context it's provided via an environment
-- variable called `ECS_CONTAINER_METADATA_URI_V4`. https://docs.aws.amazon.com/AmazonECS/latest/developerguide/task-metadata-endpoint-v4.html
-- see below for the JSON structure
ip_address INET NOT NULL,
-- under which port is the HTTP server reachable
server_port SMALLINT NOT NULL,
-- which port does the libp2p host use
peer_port SMALLINT NOT NULL,
-- a timestamp of the last heartbeat
last_heartbeat TIMESTAMPTZ,
-- a timestamp since when the node is offline (set from the scheduler if the node is unreachable, e.g., crashed)
-- but should also be set from server when shutdown gracefully.
offline_since TIMESTAMPTZ,
-- when was this node row created.
created_at TIMESTAMPTZ NOT NULL,
PRIMARY KEY (id)
);This row serves two purposes: 1) it tracks the exact configuration and dependencies of the server and 2) is used for service discovery.
The scheduler will query this table for all nodes that matches the fleet that the scheduler is after where the offline_since field is null and
the last_heartbeat is not null. In AWS, we're using VPC peering between different AWS regions, so we can use private IP addresses for connectivity between the scheduler, servers, and database.
The server must update the node_ecs row every minute to indicate it's still alive and happy to accept requests.
To expose real time metrics about the publication and retrieval performance the Go server exposes a few prometheus metrics:
metric: parsec_durations
type: summary
buckets: 50th, 90th, 95th percentile
maxAge: 24h
labels:
type: retrieval_ttfpr | provide_duration
success: true | false
scheduler: default | optprov | fullrt
metric: parsec_http_requests_total
type: summary
buckets: 50th, 90th, 95th percentile
maxAge: 24h
labels:
method: GET | POST | ...
path: /retrieve | /provide
scheduler: default | optprov | fullrt
The server can extract the available CPU and Memory from Limits.CPU and Limits.Memory. Further,
its own private IP address is in the Networks array. Look for the first entry with NetworkMode == awsvpc
and then just take the first entry of IPv4Addresses - that's a good enough heuristic so far.
{
"DockerId": "ea32192c8553fbff06c9340478a2ff089b2bb5646fb718b4ee206641c9086d66",
"Name": "curl",
"DockerName": "ecs-curltest-24-curl-cca48e8dcadd97805600",
"Image": "111122223333.dkr.ecr.us-west-2.amazonaws.com/curltest:latest",
"ImageID": "sha256:d691691e9652791a60114e67b365688d20d19940dde7c4736ea30e660d8d3553",
"Labels": {
"com.amazonaws.ecs.cluster": "default",
"com.amazonaws.ecs.container-name": "curl",
"com.amazonaws.ecs.task-arn": "arn:aws:ecs:us-west-2:111122223333:task/default/8f03e41243824aea923aca126495f665",
"com.amazonaws.ecs.task-definition-family": "curltest",
"com.amazonaws.ecs.task-definition-version": "24"
},
"DesiredStatus": "RUNNING",
"KnownStatus": "RUNNING",
"Limits": {
"CPU": 10,
"Memory": 128
},
"CreatedAt": "2020-10-02T00:15:07.620912337Z",
"StartedAt": "2020-10-02T00:15:08.062559351Z",
"Type": "NORMAL",
"LogDriver": "awslogs",
"LogOptions": {
"awslogs-create-group": "true",
"awslogs-group": "/ecs/metadata",
"awslogs-region": "us-west-2",
"awslogs-stream": "ecs/curl/8f03e41243824aea923aca126495f665"
},
"ContainerARN": "arn:aws:ecs:us-west-2:111122223333:container/0206b271-b33f-47ab-86c6-a0ba208a70a9",
"Networks": [
{
"NetworkMode": "awsvpc",
"IPv4Addresses": [
"10.0.2.100"
],
"AttachmentIndex": 0,
"MACAddress": "0e:9e:32:c7:48:85",
"IPv4SubnetCIDRBlock": "10.0.2.0/24",
"PrivateDNSName": "ip-10-0-2-100.us-west-2.compute.internal",
"SubnetGatewayIpv4Address": "10.0.2.1/24"
}
]
}Feel free to dive in! Open an issue or submit PRs.
MIT © Dennis Trautwein