AWS Security Blog

CloudHSM best practices to maximize performance and avoid common configuration pitfalls

August 31, 2021: AWS KMS is replacing the term customer master key (CMK) with AWS KMS key and KMS key. The concept has not changed. To prevent breaking changes, AWS KMS is keeping some variations of this term. More info.

AWS CloudHSM provides fully-managed hardware security modules (HSMs) in the AWS Cloud. CloudHSM automates day-to-day HSM management tasks including backups, high availability, provisioning, and maintenance. You’re still responsible for all user management and application integration.

In this post, you will learn best practices to help you maximize the performance of your workload and avoid common configuration pitfalls in the following areas:

Administration of CloudHSM

The administration of CloudHSM includes those tasks necessary to correctly set up your CloudHSM cluster, and to manage your users and keys in a secure and efficient manner.

Initialize your cluster with a customer key pair

To initialize a new CloudHSM cluster, you will first create a new RSA key pair, which we will call the customer key pair. First, generate a self-signed certificate using the customer key pair. Then, you sign the cluster’s certificate by using the customer public key as described in Initialize the Cluster section in the AWS CloudHSM User Guide. The resulting signed cluster certificate, as shown in Figure 1, identifies your CloudHSM cluster as yours.

Figure 1: CloudHSM key hierarchy and customer generated keys

Figure 1: CloudHSM key hierarchy and customer generated keys

It’s important to use best practices when you generate and store the customer private key. The private key is a binding secret between you and your cluster, and cannot be rotated. We therefore recommend that you create the customer private key in an offline HSM and store the HSM securely. Any entity (organization, person, system) that demonstrates possession of the customer private key will be considered an owner of the cluster and the data it contains. In this procedure, you are using the customer private key to claim a new cluster, but in the future you could also use it to demonstrate ownership of the cluster in scenarios such as cloning and migration.

Manage your keys with crypto user (CU) accounts

The HSMs provided by CloudHSM support different types of HSM users, each with specific entitlements. Crypto users (CUs) generate, manage, and use keys. If you’ve worked with HSMs in the past, you can think of CUs as similar to partitions. However, CU accounts are more flexible. The CU that creates a key owns the key, and can share it with other CUs. The shared key can be used for operations in accordance with the key’s attributes, but the CU that the key was shared with cannot manage it – that is, they cannot delete, wrap, or re-share the key.

From a security standpoint, it is a best practice for you to have multiple CUs with different scopes. For example, you can have different CUs for different classes of keys. As another example, you can have one CU account to create keys, and then share these keys with one or more CU accounts that your application leverages to utilize keys. You can also have multiple shared CU accounts, to simplify rotation of credentials in production applications.

Warning: You should be careful when deleting CU accounts. If the owner CU account for a key is deleted, the key can no longer be used. You can use the cloudhsm_mgmt_util tool command findAllKeys to identify which keys are owned by a specified CU. You should rotate these keys before deleting a CU. As part of your key generation and rotation scheme, consider using labels to identify current and legacy keys.

Manage your cluster by using crypto officer (CO) accounts

Crypto officers (COs) can perform user management operations including change password, create user, and delete user. COs can also set and modify cluster policies.

Important: When you add or remove a user, or change a password, it’s important to ensure that you connect to all the HSMs in a cluster, to keep them synchronized and avoid inconsistencies that can result in errors. It is a best practice to use the Configure tool with the –m option to refresh the cluster configuration file before making mutating changes to the cluster. This helps to ensure that all active HSMs in the cluster are properly updated, and prevents the cluster from becoming desynchronized. You can learn more about safe management of your cluster in the blog post Understanding AWS CloudHSM Cluster Synchronization. You can verify that all HSMs in the cluster have been added by checking the /opt/cloudhsm/etc/cloudhsm_mgmt_util.cfg file.

After a password has been set up or updated, we strongly recommend that you keep a record in a secure location. This will help you avoid lockouts due to erroneous passwords, because clients will fail to log in to HSM instances that do not have consistent credentials. Depending on your security policy, you can use AWS Secrets Manager, specifying a KMS key created in AWS Key Management Service (KMS), to encrypt and distribute your secrets – secrets in this case being the CU credentials used by your CloudHSM clients.

Use quorum authentication

To prevent a single CO from modifying critical cluster settings, a best practice is to use quorum authentication. Quorum authentication is a mechanism that requires any operation to be authorized by a minimum number (M) of a group of N users and is therefore also known as M of N access control.

To prevent lock-outs, it’s important that you have at least two more COs than the M value you define for the quorum minimum value. This ensures that if one CO gets locked out, the others can safely reset their password. Also be careful when deleting users, because if you fall under the threshold of M, you will be unable to create new users or authorize any other operations and will lose the ability to administer your cluster.

If you do fall below the minimum quorum required (M), or if all of your COs end up in a locked-out state, you can revert to a previously known good state by restoring from a backup to a new cluster. CloudHSM automatically creates at least one backup every 24 hours. Backups are event-driven. Adding or removing HSMs will trigger additional backups.


CloudHSM is a fully managed service, but it is deployed within the context of an Amazon Virtual Private Cloud (Amazon VPC). This means there are aspects of the CloudHSM service configuration that are under your control, and your choices can positively impact the resilience of your solutions built using CloudHSM. The following sections describe the best practices that can make a difference when things don’t go as expected.

Use multiple HSMs and Availability Zones to optimize resilience

When you’re optimizing a cluster for high availability, one of the aspects you have control of is the number of HSMs in the cluster and the Availability Zones (AZs) where the HSMs get deployed. An AZ is one or more discrete data centers with redundant power, networking, and connectivity in an AWS Region, which can be formed of multiple physical buildings, and have different risk profiles between them. Most of the AWS Regions have three Availability Zones, and some have as many as six.

AWS recommends placing at least two HSMs in the cluster, deployed in different AZs, to optimize data loss resilience and improve the uptime in case an individual HSM fails. As your workloads grow, you may want to add extra capacity. In that case, it is a best practice to spread your new HSMs across different AZs to keep improving your resistance to failure. Figure 2 shows an example CloudHSM architecture using multiple AZs.

Figure 2: CloudHSM architecture using multiple AZs

Figure 2: CloudHSM architecture using multiple AZs

When you create a cluster in a Region, it’s a best practice to include subnets from every available AZ of that Region. This is important, because after the cluster is created, you cannot add additional subnets to it. In some Regions, such as Northern Virginia (us-east-1), CloudHSM is not yet available in all AZs at the time of writing. However, you should still include subnets from every AZ, even if CloudHSM is currently not available in that AZ, to allow your cluster to use those additional AZs if they become available.

Increase your resiliency with cross-Region backups

If your threat model involves a failure of the Region itself, there are steps you can take to prepare. First, periodically create copies of the cluster backup in the target Region. You can see the blog post How to clone an AWS CloudHSM cluster across regions to find an extensive description of how to create copies and deploy a clone of an active CloudHSM cluster.

As part of your change management process, you should keep copies of important files, such as the files stored in /opt/cloudhsm/etc/. If you customize the certificates that you use to establish communication with your HSM, you should back up those certificates as well. Additionally, you can use configuration scripts with the AWS Systems Manager Run Command to set up two or more client instances that use exactly the same configuration in different Regions.

The managed backup retention feature in CloudHSM automatically deletes out-of-date backups for an active cluster. However, because backups that you copy across Regions are not associated with an active cluster, they are not in scope of managed backup retention and you must delete out-of-date backups yourself. Backups are secure and contain all users, policies, passwords, certificates and keys for your HSM, so it’s important to delete older backups when you rotate passwords, delete a user, or retire keys. This ensures that you cannot accidentally bring older data back to life by creating a new cluster that uses outdated backups.

The following script shows you how to delete all backups older than a certain point in time. You can also download the script from S3.

#!/usr/bin/env python

# Copyright, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: MIT-0
# Reference Links:

import boto3, time, datetime, re, argparse, typedate, json

def main():
    bkparser = argparse.ArgumentParser(prog='backdel',
                                    usage='%(prog)s [-h] --region --clusterID [--timestamp] [--timezone] [--deleteall] [--dryrun]',
                                    description='Deletes CloudHSMv2 backups from a given point in time\n')
                    help='region where the backups are stored',
                    help='CloudHSMv2 cluster_id for which you want to delete backups',
                    help="Enter the timestamp to filter the backups that should be deleted:\n   Backups older than the timestamp will be deleted.\n  Timestamp ('MM/DD/YY', 'MM/DD/YYYY' or 'MM/DD/YYYY HH:mm')",
                    help="Enter the timezone to adjust the timestamp.\n Example arguments:\n --timezone '-0200' , --timezone '05:00' , --timezone GMT #If the pytz module has been installed  ",
                    help="Set this flag to simulate the deletion",
                    help="Set this flag to delete all the back ups for the specified cluster",
    args = bkparser.parse_args()
    client = boto3.client('cloudhsmv2', args.region)
    cluster_id = args.clusterID 
    timestamp_str = args.timestamp 
    timezone = args.timezone
    dry_true = args.dryrun
    delall_true = args.deleteall
    delete_all_backups_before(client, cluster_id, timestamp_str, timezone, dry_true, delall_true)

def delete_all_backups_before(client, cluster_id, timestamp_str, timezone, dry_true, delall_true, max_results=25):
    timestamp_datetime = None
    if delall_true == True and not timestamp_str:
        print("\nAll backups will be deleted...\n")
    elif delall_true == True and timestamp_str:
        print("\nUse of incompatible instructions: --timestamp  and --deleteall cannot be used in the same invocation\n")
    elif not timestamp_str :
        print("\nParameter missing: --timestamp must be defined\n")
    else :
        # Valid formats: 'MM/DD/YY', 'MM/DD/YYYY' or 'MM/DD/YYYY HH:mm'
        if re.match(r'^\d\d/\d\d/\d\d\d\d \d\d:\d\d$', timestamp_str):
                timestamp_datetime = datetime.datetime.strptime(timestamp_str, "%m/%d/%Y %H:%M")
            except Exception as e:
                print("Exception: %s" % str(e))
        elif re.match(r'^\d\d/\d\d/\d\d\d\d$', timestamp_str):
                timestamp_datetime = datetime.datetime.strptime(timestamp_str, "%m/%d/%Y")
            except Exception as e:
                print("Exception: %s" % str(e))
        elif re.match(r'^\d\d/\d\d/\d\d$', timestamp_str):
                timestamp_datetime = datetime.datetime.strptime(timestamp_str, "%m/%d/%y")
            except Exception as e:
                print("Exception: %s" % str(e))
            print("The format of the specified timestamp is not supported by this script. Aborting...")

        print("Backups older than %s will be deleted...\n" % timestamp_str)

        response = client.describe_backups(MaxResults=max_results, Filters={"clusterIds": [cluster_id]}, SortAscending=True)
    except Exception as e:
        print("DescribeBackups failed due to exception: %s" % str(e))

    failed_deletions = []
    while True:
        if 'Backups' in response.keys() and len(response['Backups']) > 0:
            for backup in response['Backups']:
                if timestamp_str and not delall_true:
                    if timezone != None:
                        timestamp_datetime = timestamp_datetime.replace(tzinfo=timezone)
                        timestamp_datetime = timestamp_datetime.replace(tzinfo=backup['CreateTimestamp'].tzinfo)

                    if backup['CreateTimestamp'] > timestamp_datetime:

                print("Deleting backup %s whose creation timestamp is %s:" % (backup['BackupId'], backup['CreateTimestamp']))
                    if not dry_true :
                        delete_backup_response = client.delete_backup(BackupId=backup['BackupId'])
                except Exception as e:
                    print("DeleteBackup failed due to exception: %s" % str(e))
                print("Sleeping for 1 second to avoid throttling. \n")

        if 'NextToken' in response.keys():
                response = client.describe_backups(MaxResults=max_results, Filters={"clusterIds": [cluster_id]}, SortAscending=True, NextToken=response['NextToken'])
            except Exception as e:
                print("DescribeBackups failed due to exception: %s" % str(e))

    if len(failed_deletions) > 0:
        print("FAILED backup deletions: " + failed_deletions)

if __name__== "__main__":

Use Amazon VPC security features to control access to your cluster

Because each cluster is deployed inside an Amazon VPC, you should use the familiar controls of Amazon VPC security groups and network access control lists (network ACLs) to limit what instances are allowed to communicate with your cluster. Even though the CloudHSM cluster itself is protected in depth by your login credentials, Amazon VPC offers a useful first line of defense. Because it’s unlikely that you need your communications ports to be reachable from the public internet, it’s a best practice to take advantage of the Amazon VPC security features.

Managing PKI root keys

A common use case for CloudHSM is setting up public key infrastructure (PKI). The root key for PKI is a long-lived key which forms the basis for certificate hierarchies and worker keys. The worker keys are the private portion of the end-entity certificates and are meant for routine rotation, while root PKI keys are generally fixed. As a characteristic, these keys are infrequently used, with very long validity periods that are often measured in decades. Because of this, it is a best practice to not rely solely on CloudHSM to generate and store your root private key. Instead, you should generate and store the root key in an offline HSM (this is frequently referred to as an offline root) and periodically generate intermediate signing key pairs on CloudHSM.

If you decide to store and use the root key pair with CloudHSM, you should take precautions. You can either create the key in an offline HSM and import it into CloudHSM for use, or generate the key in CloudHSM and wrap it out to an offline HSM. Either way, you should always have a copy of the key, usable independently of CloudHSM, in an offline vault. This helps to protect your trust infrastructure against forgotten CloudHSM credentials, lost application code, changing technology, and other such scenarios.

Optimize performance by managing your cluster size

It is important to size your cluster correctly, so that you can maintain its performance at the desired level. You should measure throughput rather than latency, and keep in mind that parallelizing transactions is the key to getting the most performance out of your HSM. You can maximize how efficiently you use your HSM by following these best practices:

  1. Use threading at 50-100 threads per application. The impact of network round-trip delays is magnified if you serialize each operation. The exception to this rule is generating persistent keys – these are serialized on the HSM to ensure consistent state, and so parallelizing these will yield limited benefit.
  2. Use sufficient resources for your CloudHSM client. The CloudHSM client handles all load balancing, failover, and high availability tasks as your application transacts with your HSM cluster. You should ensure that the CloudHSM client has enough computational resources so that the client itself doesn’t become your performance bottleneck. Specifically, do not use resource-limited instances such as t.nano or t.micro instances to run the client. To learn more, see the Amazon Elastic Compute Cloud (EC2) instance types online documentation.
  3. Use cryptographically accelerated commands. There are two types of HSM commands: management commands (such as looking up a key based on its attributes) and cryptographically accelerated commands (such as operating on a key with a known key handle). You should rely on cryptographically accelerated commands as much as possible for latency-sensitive operations. As one example, you can cache the key handles for frequently used keys or do it per application run, rather than looking up a key handle each time. As another example, you can leave frequently used keys on the HSM, rather than unwrapping or importing them prior to each use.
  4. Authenticate once per session. Pay close attention to session logins. Your individual CloudHSM client should create just one session per execution, which is authenticated using the credentials of one cryptographic user. There’s no need to reauthenticate the session for every cryptographic operation.
  5. Use the PKCS #11 library. If performance is critical for your application and you can choose from the multiple software libraries to integrate with your CloudHSM cluster, give preference to PKCS #11, as it tends to give an edge on speed.
  6. Use token keys. For workloads with a limited number of keys, and for which high throughput is required, use token keys. When you create or import a key as a token key, it is available in all the HSMs in the cluster. However, when it is created as a session key with the “-sess” option, it only exists in the context of a single HSM.

After you maximize throughput by using these best practices, you can add HSMs to your cluster for additional throughput. Other reasons to add HSMs to your cluster include if you hit audit log buffering limits while rapidly generating or importing and then deleting keys, or if you run out of capacity to create more session keys.

Error handling

Occasionally, an HSM may fail or lose connectivity during a cryptographic operation. The CloudHSM client does not automatically retry failed operations because it’s not state-aware. It’s a best practice for you to retry as needed by handling retries in your application code. Before retrying, you may want to ensure that your CloudHSM client is still running, that your instance has connectivity, and that your session is still logged in (if you are using explicit login). For an overview of the considerations for retries, see the Amazon Builders’ Library article Timeouts, retries, and backoff with jitter.


In this post, we’ve outlined a set of best practices for the use of CloudHSM, whether you want to improve the performance and durability of the solution, or implement robust access control.

To get started building and applying these best practices, a great way is to look at the AWS samples we have published on GitHub for the Java Cryptography Extension (JCE) and for the Public-Key Cryptography Standards number 11 (PKCS11).

If you have feedback about this blog post, submit comments in the Comments session below. You can also start a new thread on the AWS CloudHSM forum to get answers from the community.

Want more AWS Security how-to content, news, and feature announcements? Follow us on Twitter.


Esteban Hernández

Esteban is a Specialist Solutions Architect for Security & Compliance at AWS where he works with customers to create secure and robust architectures that help to solve business problems. He is interested in topics like Identity and Cryptography. Outside of work, he enjoys science fiction and taking new challenges like learning to sail.


Avni Rambhia

Avni is the product manager for AWS CloudHSM. As part of AWS Cryptography, she drives technologies and defines best practices that help customers build secure, reliable workloads in the AWS Cloud. Outside of work, she enjoys hiking, travel and philosophical debates with her children.