[Hotstorage-chairs] [HotStorage 2026] Rejected, but invited as poster, submission #63 "HORIZON: A Read-Efficient Firmware for..."

[HotStorage 2026 HotCRP]

Dear authors,

The program committee for the 2026 ACM Workshop on Hot Topics in Storage 
and File Systems (HotStorage 2026) regrets to inform you that your 
submission 63 was rejected to appear in the proceedings of the workshop.

However, your work incited exciting discussion among the program 
committee members, and we would like to invite you to present your work 
during the poster session. Please let us know by August 12, 2026 
(Wednesday) whether or not you accept or decline the poster invitation.

* Title: HORIZON: A Read-Efficient Firmware for DNA Storage with 
  Horizontal Layout
* Site: https://linkprotect.cudasvc.com/url?a=https%3a%2f%2fhotstorage26.hotcrp.com%2fpaper%2f63&c=E,1,3kj0XaooKAAbxwVBDl9JFOXI8OPZFV8PnOx2Tcl64D0y8Oy7GKAzhQn3YePGywdOyDK-d1a0G8P3ARc0R6LXKXvAs31_cgPosOE2qnFdVk7nnCucoCgH&typo=1

23 of 78 submissions were accepted.

Visit the submission site for reviews, comments, and related 
information. Reviews and comments are also included below.

Contact chairs26 at hotstorage.org with any questions or concerns.

Sincerely,
Young-ri Choi and Bryan S. Kim
HotStorage 2026 Program Co-Chairs

Review #63A
===========================================================================

Overall merit
-------------
2. Weak reject

Reviewer expertise
------------------
3. Knowledgeable

Paper summary
-------------
The paper builds on the model of a DNA block device LiqSD that was introduced earlier (FAST 2025), and looks at how to map physical blocks to primers, where a primer is the "reading" unit. Since the reading unit contains many blocks, some of which are not needed, this clearly results in  read amplification. To overcome this, the paper suggests to use derived frequency information on blocks, as well as balancing the blocks across the primers (namely the reading units), to layout the physical blocks as a "horizontal distribution layout". 

The paper proposes an algorithm for the "horizontal layout" which considers both frequency information on the blocks, as well as current "occupancy level" on each primer.  Primers are divided as "retrieval units" based on "temperature".  Occupancy and temperature  guide the distribution. 

The algorithm is evaluated on the simulator of LiqSD and uses SNIA IOTTA traces and MSR traces, as well as and filesystem-based synthetic block traces. None of these are real deep archival traces. Simulations vary the "temperature-window" (the window at which frequencies are collected) as a parameter for the algorithm. As expected, simulations show that the horizontal distribution layout reduces read amplification compared to the vanilla LiqSD.  

The paper is well written and clear. It strongly builds on the FAST 2025 paper. However, my two major objections are:
- the model of a "DNA block-device firmware" and the traces used for evaluation are not realistic. DNA, if it will be successful, will be used for deep archives and not as a block device for inline storage.  
- incremental - the contribution is straightforward compared to LiqSD (see below).

Comments for authors
--------------------
Thank you for submitting your paper to HotStorage. 
The paper is clear and well written. 

Below are some concrete comments:
-   I believe that the paradigm of a "DNA block-device firmware " does not make sense in the DNA world. Although we are away from understanding how DNA storage will evolve, it is very clear that it will not replace an inline storage such as a block device due to cost and unrealistic latency. 
- The suggested addition proposed in this paper are  almost trivial to storage designers. The devised algorithm did not take into account the inherent difficulties of DNA storage. The only exception being the long "buffer time" that data remains in the buffer due to very slow "write". All considerations of distributing blocks across the reading units, adding frequency in order to reduce read amplification, are pretty straightforward in regular storage systems. 
- the idea of collecting data about frequencies before "flushing" the data  into the DNA  because write is so slow is nice.


* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *


Review #63B
===========================================================================

Overall merit
-------------
3. Weak accept

Reviewer expertise
------------------
4. Expert

Paper summary
-------------
This paper presents HORIZON, a firmware allocation policy for DNA block devices that aims to reduce read amplification through horizontal data placement, write-buffer activity classification, and primer-pair temperature tracking. The problem is timely and relevant, and the idea of improving DNA storage efficiency at the firmware layer is interesting. The proposed scheme is simple to understand, requires no changes to the biochemical layer, and the evaluation demonstrates substantial reductions in read amplification compared to LiqSD.

Comments for authors
--------------------
My primary concern is that the system model does not appear to reflect how DNA storage is currently expected to be deployed. The proposed design assumes an online block device with a continuously filling write buffer, incremental writes, and sufficient time to observe future accesses before data are committed to DNA. In contrast, current DNA storage systems are predominantly intended for archival storage, where large datasets are encoded and synthesized in bulk, with little or no rewriting. Under such a workflow, the entire dataset is typically available before synthesis, making an offline optimization of data placement across primer pairs more natural than an online firmware allocation policy. The paper should better justify why the proposed block-device model is the appropriate target for practical DNA storage systems and discuss the scenarios in which such a firmware would actually be deployed.

My second concern is that the novelty is somewhat limited. The main techniques of hot/cold data separation, temperature-aware placement, occupancy balancing, and buffering are well established ideas in storage systems and SSD firmware. While adapting these concepts to DNA storage is worthwhile, the paper should better explain what new challenges arise specifically from DNA storage and why existing allocation strategies cannot be applied directly.

Finally, the experimental evaluation could be strengthened. Most of the reported improvement appears to stem from the horizontal placement itself, while the additional activity-aware mechanisms provide more modest gains. A more detailed ablation would help quantify the contribution of each component. In addition, the evaluation reports only read amplification. Practical considerations such as metadata overhead, allocation complexity, memory requirements, and firmware overhead are not discussed, making it difficult to assess the overall cost-benefit tradeoff.

Overall, I believe the paper addresses an interesting problem, but the assumptions underlying the proposed firmware model, together with the limited novelty beyond existing storage allocation techniques and the incomplete evaluation, leave me unconvinced that the work is sufficiently mature in its current form.


* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *


Review #63C
===========================================================================

Overall merit
-------------
3. Weak accept

Reviewer expertise
------------------
2. Some familiarity

Paper summary
-------------
This paper presents HORIZON, a firmware allocation policy for reducing read amplification in DNA block storage. Its central observation is that sequentially filling primer-pair groups causes logical reads to retrieve large quantities of unrelated DNA. HORIZON instead distributes blocks horizontally across primer pairs and supplements occupancy balancing with buffered-read activity classification and sliding-window primer-pair temperature tracking. Trace-driven simulation shows consistently lower read amplification than LiqSD, with a reported 47.8× geometric-mean reduction.

Comments for authors
--------------------
The paper tackles a practical, well-framed problem that is explained clearly in the paper, and the results also show consistent benefit over the baseline. However, I have a few concerns.
Looking at Fig 8, the majority of the gains come from using horizontal, where the full results are very close to horizontal-only. More clarification and ablations to show the contributions from other components (temperature, activity classification) to the main system would be beneficial.
Along with my above comment, Fig 8 seems to be important enough to be one of the major results and should be included in the main evaluation.
Through the extended motivation, the read amplification is connected to the retrieval latency and cost, but I do not see any results that support this claim.
Extremely low values of read amplification in fig 5(a) and fig 8 for the MSR proxy seem to suggest that the system is working almost ideally. Out of curiosity, what contributions from HORIZONTAL / or dataset characteristics enable this huge benefit?


* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *


Review #63D
===========================================================================

Overall merit
-------------
2. Weak reject

Reviewer expertise
------------------
2. Some familiarity

Paper summary
-------------
The paper addresses the critical issue of Read Amplification (RA) in random-access DNA block storage devices. Due to the biochemical asymmetry between DNA writes (strand granularity) and reads (primer-pair group granularity), existing sequential state-of-the-art layout policies (e.g., LiqSD) cause heavy co-location of frequently and infrequently accessed blocks within the same retrieval units. This co-location severely drives up retrieval latency and operational cost.
To solve this, the authors propose HORIZON, a firmware allocation distributes newly written data across a wide range of primer-pair groups in a round-robin fashion instead of stacking them vertically in a single unit. It leverages an electronic write buffer to classify block hotness based on early read behaviors. The authors evaluate HORIZON against real-world block traces and synthetic filesystem workloads, demonstrating a geometric-mean reduction of 47.8× in read amplification.

Comments for authors
--------------------
Strength:

. The paper identifies and formalizes the fine-grained write vs. coarse-grained read mismatch inherent to physical DNA memory, mapping it onto a practical computer engineering abstraction (4 KiB block).

. By modulating data placement solely at the firmware level, HORIZON can provide good implementation viability.

Weaknesses:

The proposed HORIZON firmware achieves an impressive reduction in Read Amplification (RA) by utilizing a horizontal round-robin layout. However, the reviewer has a serious concern regarding the Garbage Collection (GC) efficiency and physical data fragmentation. In conventional sequential allocation (like LiqSD), block updates with high temporal locality are naturally co-located within the same primer pairs or tubes, which minimizes live-block relocation overhead during erasing (as a tube is the minimum erase unit). 

. The authors need to provide quantitative evaluation data regarding the GC overhead. Specifically, how much does HORIZON increase the number of live strands that need to be copied/relocated during a device-wide garbage collection compared to LiqSD? 

. Since DNA synthesis is extremely expensive and slow, any increase in write/relocation operations caused by inefficient GC could easily offset the gains obtained from reduced read amplification. Please discuss this trade-off in detail.


Comment @A1 by Reviewer B
---------------------------------------------------------------------------
PC Discussion Summary

The reviewers agreed that the paper addresses an interesting and timely problem by proposing a firmware-level approach for reducing read amplification in DNA block-device storage. The idea of using horizontal placement together with activity-aware allocation is well motivated, the paper is clearly written, and the evaluation demonstrates significant improvements over the LiqSD baseline.

The main concern raised during the discussion was the underlying system model. Several reviewers questioned whether the proposed online block-device abstraction, with incremental writes and a write buffer that observes future accesses before data are committed to DNA, is representative of how DNA storage is expected to be used in practice. Current DNA storage systems are primarily designed for archival storage, where large datasets are synthesized in bulk with little or no rewriting. Under this model, an offline optimization of data placement may be more natural than an online firmware allocation policy. The paper would therefore benefit from a clearer justification of the intended deployment scenario and why the proposed model is appropriate.

Another common concern was that much of the proposed approach builds on well-known storage techniques, such as hot/cold data separation, occupancy balancing, and activity-aware placement. While adapting these ideas to DNA storage is valuable, the reviewers felt that the paper should better highlight the new challenges introduced by DNA storage and explain what fundamentally differentiates the proposed approach from existing storage allocation methods.

Finally, the reviewers felt that the evaluation could be strengthened. In particular, it would be useful to better isolate the contribution of each component through additional ablation studies, and to evaluate practical aspects such as metadata overhead, firmware complexity, and the impact on garbage collection. Several reviewers also noted that the very large improvements reported for some workloads deserve additional explanation.

Overall, the reviewers found the topic interesting and believe it has the potential to stimulate future work. However, the concerns regarding the applicability of the system model, the incremental nature of the contribution, and the limited evaluation led to the decision not to accept the paper in its current form.